JPH0830581A - Method for predicting quantity of demand - Google Patents

Abstract

PURPOSE:To predict the quantity of daily demand with high predicting accuracy by inputting weather, maximum temperature and shinning hours on a current day, a record of the quantity of daily demand on a recent day with similar weather, etc., outputting the quantity of daily demand on the current day and predicting the quantity of daily demand. CONSTITUTION:At the time of learning, weather, maximum temperature and shinning hours on a current day and maximum temperature and shinning hours on a day the most recent to the current day and having similar weather which are obtained from a weather record data file 7 normalized by a normalizing device 10 and a daily water distribution record on the recent day having similar weather obtained from a water distribution record data file 6 are inputted as I/O data to a neural circuit network 11 and the quantity of daily water distribution on the day is outputted. The record values of the I/O data are set up as learning data for a certain fixed period and the network 11 is learned so that the output of the network 11 coincides with a recorded value of the quantity of daily water distribution from the file 6 normalized by the device 10. Thus the quantity of daily demand on a predicted day is predicated by non-linear approximation.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention has a non-linear relationship between a prediction target and factors affecting the prediction target.
For example, the present invention relates to a demand amount forecasting method for forecasting a distribution amount of water supply, a power demand forecast in the power field, and the like by using a learning function of a neural network.

[0002]

2. Description of the Related Art As a demand forecasting method for forecasting demand on the basis of past demand actual data of a forecast target, considering a factor affecting the forecast target, a direct fitting method, a linear approximation method, etc. , Various methods have been proposed. For example, in the case of predicting the distribution amount for each distribution reservoir by considering the weather, day of the week, etc. from the past tap water distribution amount actual data by the direct fitting method, the past demand amount actual data, , Cloudy, rain, 5 kinds of day of the week, Tuesday-Thursday, Friday, Saturday, Sundays and holidays, maximum temperature of 20 ℃ or below, 20 ℃ -30 ℃, 30 ℃
The classes are classified into the above-mentioned three kinds of 45 kinds in total, and the predicted value is obtained by averaging the data of the latest predetermined period belonging to the same class as the weather, day of the week, and maximum temperature of the predicted day, for example, the past 3 days.

Further, in the case of predicting the water distribution amount for each distribution reservoir from the past tap water distribution amount actual data by the linear approximation method, the weather, day of the week, maximum temperature, water distribution amount, etc. First, the standardized water distribution Q _S is calculated by the following formula (1).

Q _S = Q _r / (K ₁ · K ₂ · K ₃ · K ₄ ) ··· (1) Q _S : Standardized water distribution Q _r : Actual water distribution K ₁ : Actual weather coefficient K ₂ : Actual day of the week coefficient K ₃ : Actual temperature coefficient K ₄ : Correction coefficient

Here, the actual weather coefficient K ₁ is derived from a constant table, and the actual temperature coefficient K ₃ is calculated by the following equation (2).

[0006]

[Equation 1]

The standardized water distribution Q _S has an upper limit and a lower limit given by the following equation (3).

0.95Q _a ≦ Q _S ≦ 1.05Q _a (3) Q _a : Moving average of standardized water distribution for the past 6 days until the previous day

Next, based on the moving average for the past six days, for example, the day before the forecasted day of the standardized water distribution, the following equation (4)
To calculate the predicted water distribution Q _P.

Q _P = K ₁ · K ₂ · K ₃ · K ₄ · Q _a ··· (4) Q _P : Predicted water distribution Q _a : Past standardized water distribution up to the day before, for example 6 days Moving average K ₁ : Forecast weather coefficient K ₂ : Forecast day of the week coefficient K ₃ : Forecast temperature coefficient K ₄ : Correction coefficient

[0012] Note that as a document describing a technique relating to a demand forecasting method which has been conventionally used, for example, Japanese Patent Laid-Open No. 5-119808 and Japanese Patent Laid-Open No. 4-3305.
No. 02, No. 5-17976, No. 4-17976.
73332, JP-A-4-195603, and the like.

[0012]

Since the conventional demand forecasting method is configured as described above, the forecasting is basically performed based on the linear approximation such as the direct fitting method or the linear approximation method. Therefore, there is a problem in that the prediction accuracy cannot be improved when the input / output relationship is non-linear like the prediction of the tap water distribution.

The present invention has been made in order to solve the above problems, and an object thereof is to obtain a demand quantity forecasting method capable of accurately forecasting demand quantity even if the input / output relationship is non-linear. And

[0014]

The demand forecast method according to the present invention is a weather forecast, a maximum temperature, and a sunshine duration of the day, and a daily demand record, a maximum temperature, and a sunshine duration of the latest weather day. Is input, the daily demand on the day is output, the input / output relationship is learned by the neural network, and the daily demand on the forecast day is predicted.

The demand forecasting method according to the second aspect of the present invention uses the weather, maximum temperature, and sunshine duration of the day, and the actual daily demand volume, weather, maximum temperature, and sunshine duration for the past predetermined number of days as inputs. The daily demand is output as an output, and the input / output relationship is learned by the neural network to predict the daily demand.

In the demand forecasting method according to the present invention, the neural network is prepared for each class obtained by classifying past daily demand actual data into weather and maximum temperature,
For the class that includes the forecast day, the maximum temperature and sunshine hours for the day, and the actual daily demand for the same class day, maximum temperature,
With the sunshine duration as input and the daily demand as output, the input / output relationship is learned by each neural network to predict the daily demand.

In the demand forecasting method according to the invention of claim 4, the daily demand actual data is standardized in advance by linear approximation, and the weather, the maximum temperature, the sunshine time of the day and the most recent day of the same weather are standardized. The actual daily demand, the maximum temperature, and the sunshine duration are input, and the standardized daily demand for the day is output.The neural network learns the input / output relationship, and the standardized daily demand for the forecasted day is calculated. It predicts and predicts daily demand from the predicted value.

The demand forecasting method according to the invention of claim 5 obtains the average of the daily demand record and the maximum temperature for each class obtained by classifying the past daily demand record data into the weather and the maximum temperature. The weather and average maximum temperature of the class are input, the average daily demand of the class is output, and the input / output relationship is learned by the neural network to predict the daily demand.

The demand forecasting method according to the invention of claim 6 is: the difference in weather between the current day and the day before, the difference in maximum temperature, the difference in sunshine hours, and the difference in actual demand amount between the same day and the day before, the difference in weather, The maximum temperature difference and the sunshine duration difference are input, and the difference in daily demand between the current day and the previous day is output, and the input / output relationship is learned by the neural network, and the difference in daily demand between the current day and the previous day is predicted.

In the demand forecasting method according to the invention of claim 7, a neural network is prepared for each class obtained by classifying past daily demand actual data into weather and maximum temperature,
In the class that includes the forecast date, the maximum temperature difference and sunshine duration difference between the current day and the previous same class day, and the daily demand actual difference, maximum temperature difference, and sunshine difference between the latest same class day and the previous previous same class day With the time difference as input and the daily demand difference between the current day and the previous same class day as output, each neural network learns the input / output relationship and predicts the daily demand difference between the current day and the previous same class day. is there.

In the demand forecasting method according to the invention described in claim 8, the daily demand is forecasted by a direct fitting method before learning, and the weather, the maximum temperature, the sunshine time of the day and the most recent day of the same day are forecasted. The daily demand forecast error based on the forecast, the maximum temperature, and the sunshine time are input, and the daily demand forecast error based on the forecast on the day is output, and the input / output relationship is learned by the neural network to predict the daily demand. This is to predict the prediction error due to the method.

In the demand forecasting method according to the invention of claim 9, the daily demand is forecasted prior to learning by linear approximation, and the weather, maximum temperature, sunshine time of the day, and the most recent day The daily demand forecast error based on the prediction, the maximum temperature, and the sunshine duration are input, and the daily demand forecast error based on the forecast of the day is output, and the input / output relationship is learned by the neural network, and the daily demand forecast method is described. The prediction error due to is predicted.

In the demand forecasting method according to the invention described in claim 10, the average of the daily demand actual values for a predetermined period is obtained before learning, and the weather, maximum temperature, sunshine time, and recent similar weather of the day are obtained. The difference between the actual daily demand of the day and the average daily demand, the maximum temperature, and the sunshine time are input, and the difference between the daily demand calculated on the day and the average daily demand obtained is output, and the input / output relationship is the neural network. The difference between the daily demand amount on the current day and the average daily demand amount is predicted by learning.

The demand forecasting method according to the invention of claim 11 inputs the weather, maximum temperature, and sunshine duration of the day, and the actual demand amount, maximum temperature, and sunshine duration of each day of the same weather day, as input. Each time demand amount of is output as an output, the input / output relationship is learned by the neural network, and the time demand amount of each time zone of the prediction day is predicted.

The demand forecasting method according to the invention of claim 12 inputs the weather, the maximum temperature, and the sunshine duration of the day, and the actual demand demand, weather, maximum temperature, and sunshine duration for each of the past predetermined days. , The output of each hourly demand of the day, the neural network learns the input / output relationship, and predicts the hourly demand of each time zone.

In the demand forecasting method according to the invention as defined in claim 13, the neural network is prepared for each class in which the past hourly demanded actual data is classified into the weather and the maximum temperature, and the forecast date is set. In the included classes, the maximum temperature and sunshine hours on the day, and the actual demand for each hour on the latest class day,
The maximum temperature and the sunshine duration are input, the hourly demands of the day are output, the input / output relationships are learned by each neural network, and the hourly demands of each time zone are predicted.

According to the demand forecasting method of the invention described in claim 14, the average of the hourly demanded amount and the maximum temperature is averaged for each class obtained by classifying the past hourly demanded amount actual data into the weather and the maximum temperature. The weather demand of the class and the average maximum temperature are input, the output of each hourly average demand of the class, the neural network learns the input / output relationship, and the time demand of each time zone is predicted. is there.

According to the demand forecasting method of the invention described in claim 15, the difference in weather, the difference in maximum temperature, the difference in sunshine duration on the day and the day before, and the difference in the actual demand amount between the most recent weather day and the previous day,
Input the weather difference, the maximum temperature difference, and the sunshine duration difference, and output the hourly demand difference between the current day and the previous day, and let the neural network learn the input / output relationship, and predict the day of the day and the day before that. It predicts the difference in the amount of time demand.

In the demand forecasting method according to the invention as set forth in claim 16, a neural network is prepared for each class obtained by classifying the past hourly demand forecast data by weather and maximum temperature, and the forecast date is set. For the included classes, the maximum temperature difference and sunshine duration difference between the current day and the previous same class day, and the actual demand difference, maximum temperature difference, and sunshine duration difference between the latest same class day and the previous same class day of the relevant recent same class day It is used as an input, and the output of the difference in hourly demand between the same day and the previous same-class day is used to learn the input / output relationship in each neural network to predict the difference in hourly demand between the current day and the previous same-class day. is there.

The demand forecasting method according to the invention of claim 17 predicts each hour's demand in advance before learning in the same manner as the direct fitting method, and determines the weather, maximum temperature, and sunshine duration on the day. Each hour demand amount prediction error by the forecast of the most recent weather day, maximum temperature, sunshine hours are input, and each hour demand amount forecast error by the forecast of the day is output,
The neural network is made to learn the input / output relationship, and the prediction error of the time demand amount in each time zone by the prediction is predicted.

In the demand forecasting method according to the eighteenth aspect of the present invention, the average of the actual demand values for each hour in a predetermined period is obtained in advance before learning, and the weather, maximum temperature, sunshine duration, and recent The difference between the actual hourly demand and the hourly average demand on the same weather day, the maximum temperature, and the sunshine duration are input, and the difference between the hourly demand on the day and the calculated hourly average demand is output and the input The output relationship is learned by the neural network, and the difference between the time demand amount and the time average demand amount in each time zone of the day is predicted.

The demand forecasting method according to the invention of claim 19 is the weather forecast, the maximum temperature and the sunshine duration of the day, and the ratio of the hourly demand to the daily demand of the most recent weather day, the maximum temperature and the sunshine duration. Input, and output the ratio of each hourly demand amount of the day to the daily demand amount, let the neural network learn the input / output relationship, and predict the hourly demand amount of each time zone based on the separately forecasted daily demand amount. Is to do.

The demand forecasting method according to the invention of claim 20 is the weather forecast, maximum temperature, sunshine duration of the day, and the ratio of the hourly demand to the daily demand during the past predetermined number of days, the weather,
The maximum temperature and sunshine hours are input, the ratio of each hourly demand to the daily demand is output, and the input / output relationship is learned by the neural network, and each time zone based on the separately predicted daily demand is calculated. It predicts the time demand.

In the demand forecasting method according to the invention of claim 21, a neural network is prepared for each class obtained by classifying past hourly demand forecast data by weather and maximum temperature, and the forecast date is set. In the included classes, the maximum temperature and sunshine hours for the day, the ratio of the hourly demand to the daily demand for the most recent day of the same class, the maximum temperature, and the sunshine hours are entered, and the daily demand for each hourly demand for the day is input. Using the ratio as an output, each neural network is made to learn the input / output relationship, and the time demand amount of each time zone is predicted based on the separately predicted daily demand amount.

The demand forecast method according to the invention of claim 22 is, for each class obtained by classifying past hour demand record data by weather and maximum temperature, with respect to the daily demand amount of each hour demand record value. Find the average of the ratio and maximum temperature,
The weather of the class and the average maximum temperature are input, and the average of the ratio of the hourly demand of the class to the daily demand is output,
The neural network is made to learn the input / output relationship, and the time demand amount of each time zone is predicted based on the separately predicted daily demand amount.

The demand forecasting method according to the invention of claim 23 is the daily demand of the weather difference, the maximum temperature difference and the sunshine duration difference of the current day and the day before, and the hourly demand result of the most recent same day and the previous day. , The difference in the weather, the difference in the maximum temperature, the difference in the lighting time are input, and the difference in the ratio of the hourly demand of the day and the previous day to the daily demand is output, and the neural network learns the input / output relationship, The hourly demand for each time zone is predicted based on the separately forecasted daily demand.

In the demand forecasting method according to the invention as set forth in claim 24, a neural network is prepared for each class obtained by classifying the past hourly demand forecast data by weather and maximum temperature, and the forecast date is set. In the included class, the maximum temperature difference and sunshine duration difference between the current day and the previous same class day, and the difference in the ratio of the hourly demand to the daily demand amount of the latest same class day and the previous previous same class day of the relevant same class day, maximum The temperature difference and sunshine duration difference are input, and the difference in the ratio of the hourly demand to the daily demand on the same day and the previous same-class day is output, and the input / output relationship is learned by each neural network, and the forecasted date is calculated separately. The forecast of hourly demand for each time zone is based on the demand.

In the demand forecasting method according to the invention of claim 25, the ratio of each hour's demand to the daily demand is forecast in advance before learning in the same manner as the direct fitting method. Input the temperature and sunshine hours, and the prediction error of the ratio of each hourly demand amount to the daily demand amount by the above-mentioned forecast of the most recent weather day, the maximum temperature, and the sunshine time, and input the daily demand amount of each hourly demand amount by the above-mentioned forecast of the day. The output is the prediction error of the ratio to the output, the neural network learns the input / output relationship, and the time demand amount of each time zone is predicted based on the separately predicted daily demand amount.

In the demand forecasting method according to the invention described in claim 26, the average of the ratio of the hourly demand actual value to the daily demand in a predetermined period is obtained in advance before learning, and the weather and maximum temperature of the day are calculated. , Sunshine hours and the difference between the ratio of the actual hourly demand amount to the daily demand amount of the most recent weather day and the calculated average value, the maximum temperature, and the sunshine hour are input, and the hourly demand amount of each hour of the day is compared to the daily demand amount. By using the difference between the ratio and the average value as an output, the neural network learns this input / output relationship and predicts the time demand amount of each time zone based on the separately predicted daily demand amount.

According to the demand forecasting method of the invention described in claim 27, the daily demand or the hourly demand is forecast while the learning parameters are sequentially changed, and the value of the learning parameter that minimizes the forecast average error is obtained. Then, the daily demand forecast and the hourly demand forecast are performed using the learning parameter value.

In the demand forecasting method according to the invention of claim 28, the learning of the neural network is performed by excluding the data of the days when the actual demand value is significantly different from the actual demand values before and after the learning data. Is to do.

[0042]

According to the demand forecasting method of the invention described in claim 1, the weather, maximum temperature, and sunshine time of the day are input, and the actual daily demand amount, maximum temperature, and sunshine time of the same weather day are input. As the output, the neural network learns the output so that it matches the actual daily demand value, thereby predicting the daily demand amount on the forecasted day by a non-linear approximation to realize highly accurate daily demand forecasting. To do.

According to the demand forecasting method of the invention described in claim 2, the weather, the maximum temperature, and the sunshine time of the day, and the daily demand record, weather, maximum temperature, and sunshine time for the past predetermined number of days are input, and the day By using the daily demand as an output, the neural network learns the output so that it matches the actual daily demand, and predicts the daily demand by a non-linear approximation on the forecasted day, resulting in highly accurate daily demand prediction. To realize.

According to the demand forecasting method of the invention described in claim 3, the past daily demand actual data is classified into the weather and the maximum temperature, and the neural network is prepared corresponding to each class. In the class that includes the forecast date, the maximum temperature and sunshine time of the day, the actual daily demand amount of the same class day, the maximum temperature, and the sunshine time are input, and the daily demand amount is output to the neural network. By learning the output so that it matches the actual daily demand value, the daily demand amount on the predicted day is predicted by a non-linear approximation to realize highly accurate daily demand prediction.

According to the demand forecasting method of the invention described in claim 4, the daily demand result data is standardized in advance by linear approximation before learning, and the weather, the maximum temperature, the sunshine time of the day, and the recent weather are the same. Input the standardized daily demand for the day, maximum temperature, and sunshine duration, and use the standardized daily demand for the day as an output so that the output matches the standardized daily demand for the neural network. By learning
A highly accurate daily demand forecast is realized by forecasting the standardized daily demand amount on a forecast day by a non-linear approximation.

According to the demand forecasting method of the invention described in claim 5, the past daily demand actual data is divided into classes according to the weather and the maximum temperature, and the daily demand actual and the maximum temperature are classified for each class. By calculating the average, inputting the weather of the class, the average maximum temperature, outputting the average daily demand of the class, and letting the neural network learn so that the output matches the actual average daily demand of the class, The daily demand is predicted by nonlinearly approximating the relationship between the weather and maximum temperature and daily demand with noise removed as much as possible to realize highly accurate daily demand forecast.

According to the demand forecasting method of the invention described in claim 6, the difference in weather between the current day and the previous day, the difference in maximum temperature, the difference in sunshine duration, and the difference in actual demand amount between the same day and the previous day, the difference in weather, Prediction is made by inputting the maximum temperature difference and the sunshine duration difference, using the difference in daily demand between the current day and the previous day as output, and letting the neural network learn so that the output matches the difference between the actual demand values for the current day and the previous day. A highly accurate daily demand forecast is realized by predicting the difference between daily demand and the previous day's demand by nonlinear approximation.

According to the demand forecasting method of the invention described in claim 7, the past daily demand record data is classified into the weather and the maximum temperature, and the neural network is prepared corresponding to each class. In the class that includes the forecast date, the maximum temperature difference and sunshine duration difference between the current day and the previous same class day, and the daily demand actual difference, maximum temperature difference, and sunshine difference between the latest same class day and the previous previous same class day By inputting the time difference and using the daily demand difference between the current day and the previous same class day as the output, each neural network learns so that its output matches the difference between the actual daily demand value of the current day and the previous same class day. , Predicts the difference between the daily demand and the daily demand of the same class day before the predicted date by non-linear approximation to realize highly accurate daily demand forecast.

According to the demand forecasting method of the invention described in claim 8, the daily demand is forecasted by a direct fitting method before learning, and the weather, the maximum temperature, the sunshine time on the day and the most recent day Input the daily demand forecast error by prediction, maximum temperature, sunshine duration, as the output of the daily demand forecast error by the direct fitting method of the day, its output to the neural network is the actual value of the daily demand forecast error by the forecast and By learning so that they coincide with each other, the prediction error by the prediction method of the prediction date is predicted by a non-linear approximation, and highly accurate daily demand forecast is realized.

According to the demand amount predicting method of the invention described in claim 9, the daily demand amount is predicted by a linear approximation method before learning, and the weather, the maximum temperature, the sunshine time of the day, and the recent day of the same weather day are calculated. Input the daily demand forecast error based on the forecast, maximum temperature, and sunshine duration, and use the daily demand forecast error based on the forecast for the current day as the output, and its output matches the actual value of the daily demand forecast error based on the forecast. By performing the learning as described above, the prediction error of the prediction date according to the prediction method is predicted by the nonlinear approximation, and highly accurate daily demand amount prediction is realized.

According to the demand forecasting method of the invention described in claim 10, the average of the daily demand actual values for a predetermined period is obtained in advance before learning, and the weather, maximum temperature, sunshine duration and the latest weather of the day are obtained. The difference between the actual daily demand of the day and the calculated average daily demand, the maximum temperature, and the sunshine duration are input, and the output of the difference between the daily demand of the day and the average daily demand is output to the neural network. By learning to match the difference between the actual daily demand value and the average daily demand amount, the difference between the daily demand amount on the forecast day and the average daily demand amount is predicted by a non-linear approximation, and highly accurate daily demand Realize volume prediction.

According to the demand forecasting method of the invention as defined in claim 11, the weather, maximum temperature, and sunshine time of the day, and the actual hourly demand amount, maximum temperature, in each time zone of the most recent weather day,
By inputting the sunshine duration and outputting each hourly demand of the day, the neural network is trained so that its output matches the actual hourly demand, thereby nonlinearly approximating the hourly demand of each forecast time zone. To realize a highly accurate time demand forecast.

According to the demand forecasting method of the invention described in claim 12, the weather, the maximum temperature, and the sunshine time of the day, and the actual demand amount, the weather, the maximum temperature, and the sunshine time for each of the past predetermined days are input. By using each hourly demand of the current day as an output and having the neural network learn so that its output matches the actual value of the hourly demand, the hourly demand of each time zone of the forecast day is predicted by a nonlinear approximation, Realize highly accurate forecast of hourly demand.

According to the demand forecasting method of the invention described in claim 13, the past hourly demanded result data is classified into weather and maximum temperature, and a neural network is prepared corresponding to each class. In the class that includes the forecast date,
The maximum temperature and sunshine hours of the day, the hourly demand results of each recent recent class day, the maximum temperature and sunshine hours are input, and the hourly demand of the day is output, and the output is the hourly demand of each neural network. By learning so that it matches the value, the time demand in each time zone of the forecast day is predicted by nonlinear approximation,
Realize highly accurate forecast of hourly demand.

According to the demand forecasting method of the invention described in claim 14, the past daily demand actual data is classified into classes according to the weather and the maximum temperature before the learning.
Calculate the average daily demand and average maximum temperature, input the weather and average maximum temperature of the class, and output each hourly average demand of the class, and the output matches the actual hourly average demand value in the neural network. By performing learning as described above, the relationship between the weather and maximum temperature with noise removed as much as possible and each time demand amount is nonlinearly approximated to predict the time demand amount in each time zone, thereby realizing highly accurate time demand amount prediction.

According to the demand forecasting method of the invention as set forth in claim 15, the difference in weather between the current day and the previous day, the difference in maximum temperature, the difference in sunshine duration, and the difference in the actual demanded amount and the difference in weather between the most recent same day and the day before the same day. By inputting the maximum temperature difference and the sunshine duration difference, using the difference of the hourly demand amount of the current day and the previous day as the output, the neural network is trained so that the output matches the difference of the actual hourly demand value of the current day and the previous day. , Predict the time demand difference between the forecast day and the previous day by non-linear approximation to realize highly accurate time demand forecast.

According to the demand forecasting method of the invention described in claim 16, the past hourly demanded result data is classified into the weather and the maximum temperature, and a neural network is prepared corresponding to each class. In the class that includes the forecast date,
Enter the maximum temperature difference and sunshine duration difference on the same day and the same day before, and the actual demand difference, maximum temperature difference, and sunshine duration difference on the same class day and the previous recent same class day for the same class day. Forecast the current day by using each time demand difference on the same class day as output and learning each neural network so that its output matches the difference between the actual time demand value on the same day and the previous day on the same class Before the day, the difference in the amount of time demand in each time zone of the same class day is predicted by non-linear approximation to realize highly accurate time demand prediction.

According to the demand forecasting method of the invention described in claim 17, the demand demand for each hour is preliminarily predicted in the same manner as the direct fitting method before learning, and the weather, maximum temperature, sunshine duration, and recent The prediction error of each hour demand by the forecast of the same weather day, the maximum temperature, and the sunshine time are input, and the prediction error of each hour demand by the forecast method of the day is output, and the output is the time demand by the forecast. By learning so as to match the actual value of the quantity prediction error, the prediction error due to the prediction of the time demand amount in each time zone of the prediction day is predicted by a non-linear approximation, and highly accurate time demand amount prediction is realized. .

In the demand forecasting method according to the eighteenth aspect of the present invention, the average of the actual demand values for each hour in a predetermined period is calculated before learning, and the weather, maximum temperature, and sunshine duration of the day are calculated.
The difference between the actual hourly demands of the most recent weather day and the calculated average hourly demands, maximum temperature, and sunshine hours are input, and the difference between the hourly demands of the day and the hourly average demands is output, and the neural circuit is output. By learning the network so that its output matches the difference between the actual hourly demand and the average hourly demand on the day, the difference between the hourly demand and the hourly average demand in each time zone of the forecast day can be nonlinearly approximated. Forecast and realize highly accurate time demand forecast.

According to the demand forecasting method of the invention described in claim 19, the weather, the maximum temperature, and the sunshine time of the day, and the ratio of the hourly demand to the daily demand of the most recent weather day, the maximum temperature, and the sunshine time are calculated. The input and the ratio of each hourly demand of the current day to the daily demand are taken as outputs, and the neural network is trained so that the output matches the ratio of the actual hourly demand of the current day to the daily demand. Using the ratio of the hourly demand amount for each time zone of the forecast day obtained by approximation to the daily demand amount, predict the hourly demand amount for each time zone from the forecast value of the daily demand amount, and make a highly accurate hourly demand forecast. To be realized.

The demand forecasting method according to the invention of claim 20 is the weather forecast, the maximum temperature and the sunshine duration of the day, and the ratio of the hourly demand to the daily demand during the past predetermined number of days, the weather, the maximum temperature and the sunshine. By inputting the time and using the ratio of each hourly demand of the day to the daily demand as output, the neural network is trained so that the output matches the ratio of the actual hourly demand of the day to the daily demand. , Using the ratio of the hourly demand amount for each time zone of the forecast day calculated by the non-linear approximation to the daily demand amount, the hourly demand amount for each time zone is predicted from the forecast value of the daily demand amount, and the highly accurate hourly demand amount Make predictions.

According to the demand forecasting method of the invention described in claim 21, the past hourly demanded result data is classified into the weather and the maximum temperature, and a neural network is prepared corresponding to each class. In the class that includes the forecast date,
Input the maximum temperature and sunshine time of the day and the ratio of hourly demand to the daily demand of the latest class day, maximum temperature and sunshine time, and output the ratio of the hourly demand to the daily demand of each day. The neural network is trained so that its output matches the ratio of the actual hourly demand value of the day to the daily demand quantity, and the daily demand quantity of the hourly demand quantity of each forecasted time zone obtained by the nonlinear approximation is calculated. To predict the hourly demand amount of each time zone from the predicted value of the daily demand amount, and realize highly accurate forecasting of the hourly demand amount.

According to the demand forecasting method of the invention described in claim 22, the past daily demand result data is classified into classes according to the weather and the maximum temperature before learning.
The average of actual daily demand and the average of maximum temperature are calculated, the weather of the class and the average maximum temperature are input, the average of the ratio of the hourly demand of the class to the daily demand is output, and the output is output to the neural network. By learning so as to match the average of the hourly demand actual value of the class with respect to the daily demand, the relationship between the weather and maximum temperature in which noise was removed as much as possible and each hourly demand was obtained by nonlinear approximation, By using the ratio of the hourly demand amount of each time zone of the forecast day to the daily demand amount, the hourly demand amount of each time zone is predicted from the predicted value of the daily demand amount to realize highly accurate hourly demand amount prediction.

According to the demand forecasting method of the invention described in claim 23, the difference in weather between the current day and the previous day, the difference in maximum temperature, the difference in sunshine duration, and the daily demand of the recent hourly demand on the same weather day and the previous day Input the difference in the ratio, the weather difference, the maximum temperature difference, the difference in lighting time, and the output of the difference in the ratio of the hourly demand of the current day and the previous day to the daily demand is output to the neural network.
The daily demand of the hourly demand amount for each time zone of the forecast day today and the forecast day obtained by the non-linear approximation is obtained by learning so that it matches with the difference in the ratio of the actual hourly demand amount of the current day and the previous day. By using the ratio to the quantity, the hourly demand quantity of each time zone is predicted from the predicted value of the daily demand quantity, and the highly accurate hourly demand quantity prediction is realized.

According to the demand forecasting method of the invention as set forth in claim 24, the past hour demand actual result data is classified into the weather and the maximum temperature, and a neural network is prepared corresponding to each class. In the class that includes the forecast date,
Maximum temperature difference and sunshine duration difference between the current day and the previous same class day, and difference in the ratio of hourly demand to the daily demand amount of the previous same class day and the previous recent same class day of the relevant recent same class day, maximum temperature difference, sunshine duration difference The output is the difference in the ratio of the hourly demand amount to the daily demand amount between the current day and the previous same-class day, and the output to each neural network is the day of the actual hourly demand value of the current day and the previous same-class day. By learning so as to match the difference in the ratio to the demand amount, using the ratio of the time demand amount of each time zone of the same class day before the forecast day and the forecast date obtained by non-linear approximation to the daily demand amount, It predicts the hourly demand for each time zone from the forecast value of daily demand and realizes highly accurate forecasting of hourly demand.

In the demand forecasting method according to the twenty-fifth aspect of the invention, the ratio of each hourly demand to the daily demand is forecast in advance in the same manner as the direct fitting method before learning, and the weather and maximum temperature of the day are calculated. Input the forecast error of the ratio of each hourly demand amount to the daily demand amount based on the forecast of the most recent weather day, the maximum temperature, and the sunshine time, to the daily demand amount of each hourly demand amount according to the forecast method of the day. Obtained by a nonlinear approximation by using the prediction error of the ratio as an output, and letting the neural network learn so that the output matches the prediction error of the ratio of the actual value of the hourly demand to the daily demand by the prediction of the day. Using the prediction error of the above-mentioned prediction of the ratio of the hourly demand amount to the daily demand amount in each time zone of the forecast day, the hourly demand amount in each time zone is predicted from the forecast value of the daily demand amount. To realize the demand amount prediction.

In the demand forecasting method according to the invention described in claim 26, the average of the ratio of the hourly demand actual value to the daily demand in a predetermined period is calculated before learning, and the weather on the day is calculated.
Enter the maximum temperature and sunshine hours, and the difference between the ratio of the actual hourly demand amount to the daily demand amount and the average ratio of the calculated hourly demand amount to the daily demand amount, the maximum temperature, and the sunshine hour on the same day. The output of the difference between the ratio of the hourly demand to the daily demand and the average of the ratio of the hourly demand to the daily demand is output to the neural network, and the output is the daily demand of the hourly demand actual value of the day. The ratio of the hourly demand to the daily demand in each time zone of the forecasted day obtained by the nonlinear approximation by learning so that the difference between the ratio to the amount and the average of the ratio to the daily demand to each hourly demand is learned. Using the difference between the average of the ratio of each hourly demand amount to the daily demand amount, the hourly demand amount of each time zone is predicted from the predicted value of the daily demand amount, and highly accurate hourly demand amount prediction is realized.

According to the demand forecasting method of the invention as set forth in claim 27, the daily demand or the hourly demand is forecast while the learning parameters are sequentially changed, and the value of the learning parameter that minimizes the forecast mean error is obtained. By performing the daily demand forecast and the hour demand forecast with the learning parameter value, highly accurate daily demand forecast and time demand forecast are realized.

In the demand forecasting method according to the invention of claim 28, learning is performed by performing learning on the neural network except for days when the actual demand value is significantly different from the actual demand values before and after the learning data. The noise contained in the data is removed to realize highly accurate daily demand forecast and hourly demand forecast.

[0070]

【Example】

Example 1. Embodiment 1 of the present invention will be described below with reference to the drawings. FIG. 1 is a conceptual diagram showing a daily water distribution forecasting system for water supply to which a demand forecasting method according to an embodiment of the invention described in claim 1 is applied. In the figure, 1 is a distribution flow rate detection device 1 for detecting the distribution amount of water in the water supply to obtain distribution record data, and 2 is a past distribution amount record data obtained from this distribution flow rate detection device 1 and a past meteorological record. data,
The daily distribution amount predicting device 3 for predicting the daily distribution amount on the basis of the weather prediction data on the day of the prediction day is a predicted value of the daily distribution amount predicted by the daily distribution amount predicting device 2. Numeral 4 is an hourly water quantity predicting device that predicts the hourly water quantity for each time zone based on this daily water quantity forecast value 3, past water quantity actual data, and past meteorological related actual data. It is a water distribution control device that controls the water distribution of water supply based on the hourly water distribution amount predicted by the water amount prediction device 4.

Further, in the daily water distribution forecasting apparatus 2, 6
Is a water distribution result data file in which the past water distribution result data is stored, 7 is a meteorological result data file in which the past meteorological result data is stored, and 8 is this water distribution result data file 6 and meteorological result Reference numeral 9 is a performance database including a data file 7, and 9 is a weather forecast data file in which the weather forecast data of the day is stored. Reference numeral 10 is a normalization device that normalizes the past water distribution result data, the past meteorological-related performance data, and the weather forecast data of the day, and 11 is the normalization device 10
Input the past water distribution result data and the past meteorological-related result data that have been normalized in to perform learning, and use the normalized past water distribution result data, the past meteorological-related result data, and the weather forecast data of the day. It is a neural network that predicts the daily distribution of water on the forecast day, and the neural network 11 uses as input and output data the weather of the day, the maximum temperature of the day, the sunshine hours of the day, and the daily distribution of the most recent weather day. Input the actual amount of water, the maximum temperature of the most recent weather day, and the sunshine hours of the most recent weather day, and output the daily water distribution on that day.

Next, the operation will be described. In the daily water distribution predicting apparatus 2, first, the past water quantity actual data obtained from the water quantity detecting unit 1 is stored in the actual water quantity data file 6 in the actual performance database 8, and the past weather and maximum temperature are recorded in the meteorological-related performance data file 7. In addition to storing meteorological-related performance data such as sunshine hours, the weather forecast data file 9 also stores forecast weather data such as expected weather, maximum temperature, and sunshine hours on the day of forecast of daily water distribution. These data are normalized by the normalizing device 10, sent to the neural network 11 as input / output data, and used for learning and prediction of daily water supply. Here, in the normalization device 10, in the input / output data of the neural network 11, the maximum temperature of the day, the sunshine time of the day, the actual amount of daily water distribution on the same weather day,
The maximum temperature on the same day, the sunshine hours on the same day, and the water distribution on the day are normalized using the following equation (5). For the weather on the day, fine = 1, cloud = 0 ．
Normalize as 5, rain = 0.

Xn = (X−Xmin) / (Xmax−Xmin) (5) Xn: normalized value of each input / output value X: each input / output value Xmin: minimum value of each input / output value -Α Xmax: maximum value of each input / output value + β α, β: correction value

At the time of learning, as the input / output data of the neural network 11, the weather of the day, the maximum temperature of the day, and the day of the day from the meteorological-related result data file 7, which are normalized as described above by the normalizing device 10, respectively. Sunshine hours,
The highest temperature of the day with the closest same day (the most recent same-day day), the sunshine duration of the most recent day of the same day, and the daily water quantity record of the most recent same-day day from the water quantity record data file 6 are input. The daily water supply is output. Then, using the actual value of the input / output data for a certain period, for example, the past three weeks as learning data, the output of the neural network 11 and the normalization device 10 normalize the distribution amount actual data file 6 The neural network 11 is learned so that the actual values of the daily water distributions match. However, in this case, only weekday data are treated as learning data.

The neural network 11 to be learned will be described below with reference to FIGS. FIG. 2 shows a neural network called a hierarchical neural network, FIG. 3 shows a recurrent neural network, and FIG. 4 shows a chaotic neural network.

In the hierarchical neural network shown in FIG. 2, the neurons of the input layer and the intermediate layer, and the neurons of the intermediate layer and the output layer are connected so that the weight of the connection is changed by learning, and each of the learning data is input to the input layer. Is input to the output layer to output the actual value of the drainage amount, and the learning operation is performed as follows. Total y _k (t) of inputs to the k-th neural circuit element k in each layer other than the input layer of the neural network
The output of the x _l of l-th neural elements l prior layer, the connection weights of the neural elements l and neural elements k and w _kl, represented by the following equation (6).

[0077]

[Equation 2]

At this time, the next time value z _k (t + 1) in the k-th neural circuit element is given by the following equation (7).

Z _k (t + 1) = f _k (y _k (t)) (7)

Here, f _k is a transfer function of the neural circuit element, and normally, a function called a sigmoid function represented by the following equation (8) is often used.

[0081]

(Equation 3)

Here, the back-propagation method, which is a learning method of the hierarchical neural network, is executed as follows. That is, the error evaluation function E _p for the p-th input / output data is t _{pj for} the j component of the teacher data for the p-th input data, and y _{pj for} the j component of the actual output of the network for the p-th input data. Then, the following equation (9) is obtained.

[0083]

[Equation 4]

At this time, the change amount of the connection weight W _ji is given by the following equation (10), where η is the learning coefficient and α is the inertia coefficient.

[0085]

(Equation 5)

In this learning, input / output data is repeatedly presented until the error evaluation function E _p has a sufficiently small value, and the weight of the connection is updated.

The operation of the recurrent neural network shown in FIG. 3 is the same as the operation of the hierarchical neural network shown in FIG. 2 with feedback from the intermediate layer to the input layer. Therefore, also as the learning method of this recurrent neural network, the above-mentioned backpropagation method can be used by ignoring the feedback connection.

Further, FIG. 4 shows only the chaotic neuron #i of the chaotic neural network, and its operation is performed as follows. That is, #i
The internal state y _i (t + 1) of the chaotic neuron of time j is the strength of synaptic connection from the chaotic neuron of #j to the chaotic neuron of #i w _ij , and the time t of the chaotic neuron of #j. Output is x _j (t), a function representing the conversion characteristic of the axon of the chaotic neuron #j is h _j , and the strength of synaptic connection from the external input I _j (t) of #j to the chaotic neuron #i is strong. Where v _ij is the height, F _{i is} the function representing the refractory, and θ _i is the threshold value, they are given by the following equation (11).
Note that k _i in this equation (11) is a parameter that satisfies 0 <k _i <1.

[0089]

(Equation 6)

Here, the output x _i (t + 1) of the chaotic neuron _{#i at time} (t + 1) is given by the following equation (12), where G _i is the output function of the chaotic neuron _#i. Will be used.

X _i (t + 1) = G _i (y _i (t + 1)) (12)

At the time of forecasting, the neural network 11 after learning in this way is predicted by the normalizing device 10 and the forecasted weather of the amount of water distribution from the weather forecast data file 9 on the day of the forecast. , The predicted maximum temperature of the day,
Predicted sunshine hours on the current day, and the forecasted date from the water distribution amount data file 6 The actual water amount on the day with the same weather (the most recent same weather day) closest to the day, and the most recent weather day from the weather-related result data file 7 The maximum daily temperature and the sunshine duration on the same weather day are input to predict the normalized daily water distribution amount on the predicted date, and denormalize it to output it as the daily water amount prediction value 3. However, even in this case, only weekdays will be treated as the forecast date.

When the water distribution control device 5 controls the water distribution of the water supply, the hourly water distribution amount predicting device 4 uses the daily water amount predicting value 3 output from the daily water amount predicting device 3 for each time. The water distribution amount for each zone is predicted, and the predicted value of the time water distribution amount is input to the water distribution control device 5 as data for water supply water distribution control.

Example 2. Here, in the first embodiment, as the input / output data of the neural network, the weather of the day, the maximum temperature of the day, the sunshine time of the day, the actual water distribution amount of the most recent weather day, the maximum temperature, and the sunshine time are input. Although the case of outputting the daily water distribution on the day was shown, as the input / output data of the neural network, the weather of the day, the maximum temperature of the day, the sunshine hours of the day, the actual daily water distribution of the past T days, the weather, the maximum The temperature and sunshine duration may be input and the daily water distribution amount may be output, and the same effect as that of the above-described embodiment is obtained. In this case, at the time of forecasting, enter the forecasted weather, forecasted maximum temperature, forecasted sunshine duration, daily water distribution results for the past T days, weather, maximum temperature, and sunshine duration into the neural network to forecast Predict water volume. The second embodiment is an embodiment of the invention described in claim 2.

Example 3. Next, a third embodiment of the present invention will be described with reference to the drawings. FIG. 5 is a conceptual diagram for explaining a demand quantity forecasting method according to an embodiment of the invention described in claim 3. In the third embodiment, before learning,
Based on the weather and the maximum temperature, the daily distribution amount actual data of the past predetermined period, for example, the past several months are classified into a plurality of classes, and the neural network is prepared for each of the classes. In the example shown, the weather is clear, cloudy, or rainy.
There are 9 neural networks 11 _{1 to} 11 ₉ corresponding to each class, with the maximum temperature being 20 ° C or less, 20 ° C to 30 ° C, and 30 ° C or more There is. Then, in containing the predicted day class, as input data input and output to neural network 11 ₁ to 11 _9, the maximum temperature of the day, sunshine hours of the day, the closest the same class the day day (recently the (Class day) actual amount of water distribution, maximum temperature of the same class day, sunshine duration of the same class day recently, and output of water distribution on the day of the same day Using the actual values of input / output data belonging to the same class before and after 8 days as learning data, the input / output relationship is trained by the neural network 11 _{1 to} 11 ₉ of the class including the prediction date. Now, if the class including the predicted date is a cloudy day at 20 ° C. or less shown by shading in FIG. 5, the shaded neural network 11 ₂ shows the same class day around 8 days in the past. The relevant 20 ℃
The actual values of input / output data belonging to the following cloudy day classes are used as learning data, and learning is performed so that the output of the neural network 11 ₂ matches the actual daily water distribution amount value.

After the learning is completed, the daily water distribution is predicted using the neural network corresponding to the class including the predicted day. Now, class prediction on the day belongs, for example, if the 20 ℃ following of cloudy days, the predicted maximum temperature of the predicted date of neural network 11 ₂ corresponding to the class, prediction daylight hours, and recently the same class Date Actual daily water distribution, maximum temperature,
Enter the sunshine duration to predict the daily water distribution on the forecast day.

Here, the conventional direct fitting method, the method of the first embodiment, and the third embodiment of the present invention are used by using the water distribution amount actual data for one year and the meteorological data in FIGS. 6, 8 and 10. The results of daily forecast of water distribution are shown below. Each of these targets weekdays only, the first 50 days are used for initial learning, and about 19 days after that.
For day 0, the prediction was made by repeating learning. In addition, the solid line in the figure shows each predicted value, and the dotted line shows each actual value. Further, FIGS. 7, 9, and 11 are explanatory diagrams in which the respective prediction errors at that time are represented by a graph, and FIG. 12 shows the graphs of the errors by these three methods in one for easy comparison. It is an explanatory view shown. In FIG. 12, the solid line shows the prediction error by the method of the first embodiment, the broken line shows the prediction error by the method of the third embodiment, and the dotted line shows the prediction error by the direct fitting method which is the conventional method. Further, FIG. 13 is an explanatory diagram in which the cumulative number of days due to the prediction error in each of the above methods is represented by a graph, and in this case also, the prediction error according to the method of the first embodiment is a solid line, and the prediction error according to the method of the third embodiment is a broken line. so,
The dotted line shows the prediction error of the conventional direct fitting method. Further, FIG. 14 is an explanatory diagram for comparing prediction errors in the respective methods, and the maximum error, the minimum error, and the average error for the respective methods are shown in a table format. As is clear from FIGS. 12 to 14, it can be seen that the method according to the first and third embodiments has higher prediction accuracy than the conventional method.

Example 4. Next, a fourth embodiment of the present invention will be described with reference to the drawings. FIG. 15 is a conceptual diagram for explaining a demand quantity forecasting method according to an embodiment of the invention described in claim 4. As shown in the figure, in the fourth embodiment, the daily distribution amount actual data is standardized in advance by the linear approximation method as described as the prior art. Then, as the input / output data of the neural network 11, the weather of the day, the maximum temperature of the day, the sunshine time of the day, the standardized daily water distribution record of the recent weather day, the maximum temperature of the recent weather day,
The neural network 11 is learned by inputting the sunshine duration of the same weather day and using the standardized daily water supply amount of the day as an output. For this learning, the actual value of the input / output data for a certain fixed period, for example, a period of about 3 weeks is used as the learning data so that the output of the neural network 11 matches the standardized actual daily water supply amount value. The neural network 11 is learned.

After the learning is completed, the neural network 11 is used to predict the standardized daily water distribution amount on the prediction day.
The forecast of this standardized daily water distribution is based on the forecasted weather, predicted maximum temperature, predicted sunshine duration, and standardized daily water distribution results of the same day on the neural network 11 after learning. It is performed by inputting the maximum temperature of the most recent weather day and the sunshine duration of the most recent weather day to the neural network 11. When the standardized daily water distribution on the forecast day is predicted, the forecasted value is inversely standardized by the above-mentioned linear approximation method to obtain the daily water forecast on the forecasted day.

Example 5. Next, a fifth embodiment of the present invention will be described. This Embodiment 5 is an embodiment of the invention described in claim 5, and prior to learning, the past daily distribution amount actual data is classified into weather and maximum temperature, and the daily distribution amount actual data of each class is calculated. Take the average and the average of the maximum temperature. For example, the daily water distribution results data for the past few months shows three weather conditions: fine, cloudy, and rain, and maximum temperature of 1
0 ° C or lower, 10 ° C to 13 ° C, 13 ° C to 16 ° C, 16 ° C to
19 ° C, 19 ° C-22 ° C, 22 ° C-25 ° C, 25 ° C-2
8 classes, 28 degrees Celsius-31 degrees Celsius, and 31 degrees Celsius or more are classified into a total of 27 classes, and the daily distribution volume actual data and the maximum temperature are averaged for the daily distribution volume actual data for each class.

Then, as the input / output data of the neural network, the weather of the class and the average maximum temperature of the class are input, and the average daily water distribution of the class is output, and the neural network is learned. For this learning, the actual values of the input / output data of all classes are used as learning data, and the neural network 11
The learning of the neural network 11 is performed so that the output of 1 matches the actual value of the average daily water distribution of the class. By using the neural network 11 after the learning, and inputting the forecast weather, the forecast maximum temperature, and the forecast sunshine time to the forecast date of the water distribution amount, the daily water amount of the forecast day is predicted.

Example 6. Next, a sixth embodiment of the present invention will be described with reference to the drawings. FIG. 16 is a conceptual diagram showing input / output data of the neural network 11 in an embodiment of the invention described in claim 6. As shown in the figure, in this sixth embodiment, as the input / output data to the neural network 11, the weather difference between the current day and the previous day, the maximum temperature difference between the current day and the previous day, the sunshine duration difference between the current day and the previous day, and the recent same weather Difference in actual water distribution between the day and the day before the same day, the difference in weather between the same day and the day before the same day, the maximum temperature difference between the same day and the day before the same day, the latest The sunshine duration difference between the same weather day and the day before the same weather day is input, and the difference in daily water distribution between the current day and the previous day is output.

For the learning of the neural network 11, the actual value of the input / output data is used as the learning data, and the neural network 11 outputs the neural network so that the output of the neural network 11 coincides with the actual difference value of the daily distribution amount of the current day and the previous day. The circuit network 11 is learned. In addition, in the neural network 11 after the end of the learning, the forecasted weather difference on the forecasted day and the forecasted day before, the forecasted maximum temperature difference, the forecasted sunshine duration difference, the forecasted most recent weather day and the most recent previous weather day's day. By inputting the actual water distribution difference, the weather difference, the maximum temperature difference, and the sunshine duration difference, the difference in the daily water distribution between the predicted day and the day before the predicted day is predicted. Then, based on this predicted value and the actual value of the daily water distribution on the day before the predicted date, the daily water distribution on the predicted day is predicted.

Example 7. Next, a seventh embodiment of the present invention will be described with reference to the drawings. FIG. 17 is a conceptual diagram showing the input / output data of the neural network in the embodiment of the invention described in claim 7. As shown, in this seventh embodiment,
Prior to learning, in the same way as in the case of Example 3, the past daily water distribution result data is classified into a plurality of classes according to the weather and the maximum temperature (in this case, the weather is 3 and the maximum temperature is 3 in total 9 classes). , And the neural network 11 ₁
To 11 ₉ are prepared. Then, in the class included in the predictive beautification, the maximum temperature difference between the current day and the previous same-class day, the sunshine duration difference between the current day and the previous same-class day, and the recent same-class as input / output data to the neural networks 11 _{1 to} 11 ₉ Day and the day before the same class day Actual water distribution difference between the day and the same class day, recent temperature difference between the same class day and the day before the same class day Maximum temperature difference between the same class day, the recent same class day and the day before the same class day The difference in sunshine hours from the day of the same class is input, and the difference in daily water distribution between the current day and the day before the same class is output.

For learning the neural network 11 ₁ (to 11 ₉ ), the actual value of this input / output data is used as learning data, and the output of the neural network 11 ₁ (to 11 ₉ ) is the same as the day before. The neural network 11 ₁ (to 11 ₉ ) is learned so as to match the actual difference in the daily water supply amount on the class day. In addition, after the completion of the learning, the neural network 11 ₁ to 11 ₉ which corresponds to the class that contains the prediction on the day, the predicted maximum temperature difference between the before the class day of the predicted day and predicted the day, predicted sunshine time difference, By inputting the actual difference in the amount of daily water distribution, the difference in maximum temperature, and the difference in sunshine duration between the same class date and the previous previous class day, the daily water distribution between the forecast day and the same day before the forecast day. Predict the difference. Then, based on this predicted value and the actual value of daily water distribution on the same class day before the predicted day, the daily water distribution on the predicted day is predicted.

Example 8. Next, an eighth embodiment of the present invention will be described with reference to the drawings. FIG. 18 is a conceptual diagram showing input / output data of the neural network 11 in the embodiment of the invention described in claim 8. As shown in the figure, in the eighth embodiment, the daily distribution amount is predicted in advance before the learning by the direct fitting method as described in the conventional art.
Then, as input / output data to / from the neural network 11, the weather of the day, the maximum temperature of the day, the sunshine time of the day, the error in forecasting the daily water distribution by the prediction method of the recent weather day, the maximum temperature of the recent weather day, the recent The sunshine duration on the same weather day is input, and the daily water distribution forecast error of the forecast method of the current day is output.

For learning of the neural network 11, the actual value of this input / output data is used as learning data, and the output of the neural network 11 matches the actual value difference of the daily water distribution prediction error by the above-mentioned prediction method. Thus, the neural network 11 is learned. In addition, in the neural network 11 after the end of this learning, the forecasted weather on the forecasted day, the forecasted maximum temperature, the forecasted sunshine time, the daily water distribution forecast error of the most recent weather day, the maximum temperature of the most recent weather day, the most recent weather day By inputting the sunshine duration of the day, the prediction error based on the prediction method on the prediction day is predicted.
Then, based on this predicted value and the predicted value according to the above-mentioned prediction method, the daily water distribution amount on the predicted day is predicted.

Example 9. Here, in the eighth embodiment, the case where the daily distribution amount is predicted in advance by the direct fitting method before learning has been described. However, the daily distribution amount may be predicted in advance before learning by the linear approximation method as described in the related art. The amount of water may be predicted, and the same effect as that of the above-described eighth embodiment is obtained. In that case, the input / output data of the neural network includes the weather of the day, the maximum temperature of the day, the sunshine time of the day, the daily water distribution prediction error by the linear approximation method of the most recent day of the same day, the most recent day of the same day. Maximum temperature of
The sunshine duration of the same weather day is input, and the daily water distribution prediction error of the linear approximation method of the day is output.

Example 10. Next, Example 10 of the present invention
Will be described with reference to FIG. FIG. 19 is a conceptual diagram showing input / output data of the neural network 11 in the embodiment of the invention described in claim 10. As shown, this Example 10
In the above, the average of the daily water distribution actual values for a predetermined period, for example, the last month, is obtained in advance before learning. Then, as input and output data to the neural network 11, the weather of the day, the maximum temperature of the day, the sunshine time of the day, the difference between the actual value of the daily water distribution and the average water distribution on the most recent day, the most recent day of the same weather, The maximum temperature and the sunshine duration on the same weather day are input, and the difference between the daily water distribution on the day and the average daily water distribution is output.

For learning of the neural network 11, the actual value of this input / output data is used as learning data, and the output of the neural network 11 is the difference between the actual value of daily water distribution on the day and the average daily water distribution. The neural network 11 is learned so that In addition, in the neural network 11 after the end of this learning, the predicted weather of the day, the predicted maximum temperature of the day, the predicted sunshine time of the day, the daily water distribution actual value of the most recent weather day, and the average daily water distribution By inputting the difference, the maximum temperature on the most recent weather day, and the sunshine duration on the most recent weather day, the difference between the daily water distribution on the forecast day and the average daily water distribution is predicted. Then, based on this predicted value and the average daily water distribution, the daily water distribution on the predicted day is predicted.

Example 11. As described above, in Examples 1 to 10, the case of predicting the daily water distribution amount on the prediction day has been described, but the hourly water distribution amount in each time zone on the prediction day may be predicted. The eleventh embodiment is an embodiment of the invention described in claim 11, and the input / output data of the neural network is the weather of the day, the maximum temperature of the day, the sunshine time of the day, and a predetermined time of the most recent weather day. The amount of hourly water distribution for each zone (in the following examples, it will be described as the amount of hourly water distribution for each time zone from 1:00 to 24:00) performance, the maximum temperature of the most recent weather day, and the sunshine hours of the most recent weather day. A total of 29 water is input, and a total of 24 hourly water distributions in each time zone of the day is output.

At the time of learning the neural network, the actual value of this input / output data is used as learning data, and the neural network is learned so that the output of the neural network matches the actual amount of water distributed over time. In addition, at the time of forecasting, forecasted weather on the forecasted day, forecasted maximum temperature on the day, forecasted sunshine time on the day, actual water distribution for each hour on the latest weather day, maximum temperature on the latest weather day, sunshine duration on the latest weather day Is input to the neural network after learning to predict the amount of water distributed in each time zone on the predicted day.

Example 12. Here, in the eleventh embodiment, as the input / output data of the neural network, the weather of the day, the maximum temperature, the sunshine time, the actual amount of water distribution in each time zone of the most recent weather day, the maximum temperature of the most recent weather day, Although the case of inputting the sunshine time is shown, as the input / output data of the neural network, instead of the actual amount of water distribution in each time zone of the latest weather day, and the maximum temperature and sunshine time of the latest weather day, It is possible to input the actual hourly water amount in each time zone from 1 o'clock to 24 o'clock in the past T days, and the weather, maximum temperature, and sunshine duration for the past T days, and the same effect as in the above-described embodiment is obtained. . In this case, at the time of forecasting, the forecasted weather, forecasted maximum temperature, forecasted sunshine time, actual amount of water distribution for each time zone of the past T days, weather for the past T days, maximum temperature, sunshine duration are input to the neural network. Then, the water distribution for each hour on the forecast day is forecast. The twelfth embodiment is an embodiment of the invention described in claim 12.

Example 13 Next, Example 13 of the present invention
Will be described. The thirteenth embodiment is an embodiment of the invention described in the thirteenth aspect. Prior to learning, for example, the hourly water distribution result data of the past several months is based on the weather and the maximum temperature, for example, fine weather, Cloudy, rainy 3
The maximum temperature is 20 ° C or lower, 20 ° C to 30 ° C, and 30 ° C or higher. Then, prepare nine neural networks corresponding to each of the classes, and in the class including the predicted date, the maximum temperature of the day, as the input / output data to / from the neural network,
Enter the total of 28 hours of sunshine on the day, hourly water distribution in each time zone of the same class day, maximum temperature of the same class day, and sunshine hours of the same class day of the last day. The output of the neural network is learned so that the output of the neural network coincides with the actual value of the hourly water distribution, for example, with respect to this input / output relationship, for example, for the data belonging to the same class around the past 8 days.

After the end of this learning, the neural network corresponding to the class to which the prediction day belongs belongs to the predicted maximum temperature on the prediction day, the predicted sunshine time, the actual amount of water distribution for each hour on the latest class day, and the recent same class. Enter the maximum day temperature and sunshine hours,
Predict the hourly water distribution in each time zone from 1:00 to 24:00 on the forecast day.

Example 14 Next, Example 14 of the present invention
Will be described. The fourteenth embodiment is an embodiment of the invention described in claim 14, and before the learning, for example, each hourly water distribution result data of the past several months is based on the weather and the maximum temperature, and the weather is fine. , Cloudy, rain, maximum temperature below 10 ℃, 10 ℃ to 13 ℃, 13 ℃
-16 ° C, 16 ° C-19 ° C, 19 ° C-22 ° C, 22 ° C-
25 classes, 25 degrees C-28 degrees C, 28 degrees C-31 degrees C, 31 degrees C or more, are classified into a total of 27 classes, and in each class, the hourly water distribution results of each hour from 1 o'clock to 24 o'clock Take the average and the maximum temperature.

Then, as the input / output data of the neural network, two of the weather of the class and the average maximum temperature of the class are input, and a total of 24 hourly average water distributions from 1:00 to 24:00 of the class is output. The actual values of the input / output data of all classes are used as learning data, and the neural network is learned so that the output of the neural network matches the actual value of each time average water distribution of the class. Then, using the neural network after this learning, the forecast weather of the water distribution forecast day,
By inputting the predicted maximum temperature, the hourly water distribution for each time zone from 1 o'clock to 24 o'clock on the predicted day is predicted.

Example 15. Next, Embodiment 15 of the present invention
Will be described. This fifteenth embodiment is one embodiment of the invention described in claim 15, and the input / output data to / from the neural network is the weather difference between the current day and the previous day, the maximum temperature difference between the current day and the previous day, and the current day and the previous day. Difference in sunshine hours, difference in actual amount of water distributed in each time zone from 1 o'clock to 24 o'clock between the last day of the same weather and the day before the last day of the same weather, weather difference between the last day of the same weather and the day before the last day of the same weather, maximum temperature A total of 30 differences and sunshine duration differences are input, and a total of 24 differences in the amount of water distributed for each time period from 1:00 to 24:00 between the current day and the previous day are output.

In learning the neural network, the actual values of the input / output data are used as learning data, and the neural network is adjusted so that the output of the neural network coincides with the actual value difference between the hourly water distribution and the previous day. Is learned. In addition, the neural network after this learning shows the predicted weather difference between the predicted day and the day before the predicted day, the predicted maximum temperature difference, the predicted sunshine duration difference, the most recent weather day of the predicted day, and the most recent weather day of the predicted day. By inputting the actual difference in the amount of water distributed in each time zone from the previous day, the difference in weather between the most recent weather day of the forecast day and the day before the most recent weather day of the forecast day, the maximum temperature difference, and the difference in sunshine duration, Predict the difference in the amount of hourly water distribution between each day and the day before the forecast day. Then, based on this predicted value and the actual value of the hourly water distribution on the day before the predicted date,
Forecast The forecast of hourly water distribution for each time of day.

Example 16. Next, Embodiment 16 of the present invention
Will be described. This sixteenth embodiment is an embodiment of the invention described in claim 16, and similarly to the case of the thirteenth embodiment, before the learning, for example, the hourly water distribution actual result data of the past several months is obtained from the weather and the maximum temperature. On the basis of, for example, weather is fine, cloudy, rain, and maximum temperature is 20 ℃.
In the following, three classes of 20 ° C. to 30 ° C. and 30 ° C. or more are classified into a total of nine classes, and nine neural networks are prepared for each class. Then, as the input / output data of each of those neural networks, the maximum temperature difference between the current day and the previous same-class day, the sunshine duration difference between the current day and the previous same-class day, the recent same-class day and the previous same-class day The total difference of the actual amount of water distributed in each time zone from 1 o'clock to 24 o'clock with the day, the maximum temperature difference between the latest same class day and the previous same class day, and the difference in sunshine hours 2
Input eight, and output a total of 24 differences in the amount of water distribution for each time period from 1 o'clock to 24 o'clock on the current day and the same class day before.

The learning of the neural network is performed on the neural network corresponding to the class including the prediction day, and the actual values of the input / output data of the past several days of the same class are used as the learning data. The learning of the neural network is performed so that the output of is in agreement with the actual difference in the amount of water distributed in each time period between the current day and the previous class day. In addition, after this learning, the neural network corresponding to the class that includes the predicted day of the day, the predicted maximum temperature difference between the predicted day of the day and the previous class day of the predicted day, the predicted sunshine duration difference, the latest same class day By inputting the actual difference in the amount of water distribution for each hour from the previous same day on the same class day, the maximum temperature difference between the latest same class day and the previous same class day on the latest same class day, and the difference in sunshine hours, Predict the hourly water supply difference between each day and the same class day before the forecast day. Then, based on this predicted value and the actual value of each hourly water distribution for the same class day before the predicted day, the hourly water distribution for each time zone on the day of the predicted day is predicted.

Example 17 Next, Example 17 of the present invention
Will be described. This embodiment 17 is an embodiment of the invention described in claim 17, and in this embodiment 17, before learning, a direct fitting method equivalent to that for the daily water distribution amount performed in embodiment 8 is used in advance. By classifying the hourly water distribution data by weather, maximum temperature, and day of the week, predict hourly water distribution from 1 o'clock to 24 o'clock as the average of the hourly water distribution actual values belonging to the same class, for example, the past 3 days. deep.
And, as the input / output data of the neural network, the weather of the day, the maximum temperature of the day, the sunshine time of the day, the prediction error of the hourly water distribution in each time zone from 1 o'clock to 24 o'clock according to the prediction method of the most recent weather day, The maximum temperature of the most recent weather day, the total of 29 sunshine hours of the most recent weather day are input, and the total 24 errors of the hourly water distribution prediction error in each time zone from 1 o'clock to 24 o'clock according to the prediction method of the day are output. To do.

In learning the neural network, the actual value of this input / output data is used as learning data, and the neural network is adjusted so that the output of the neural network matches the actual value of the temporal water distribution prediction error by the prediction method. The network is learned. Also,
After this learning is completed, the neural network predicts the forecast weather, forecast maximum temperature, forecast sunshine time on the forecast day, error in forecasting the amount of water distribution in each time zone according to the forecast method of the same weather day, and the highest of the latest weather day. By inputting the temperature and the sunshine duration, the prediction error of the hourly water distribution in each time zone of the prediction day by the prediction method is predicted. Then, based on this predicted value and the predicted value obtained by the prediction method, the hourly water distribution for each time zone on the predicted day is predicted.

Example 18. Next, Embodiment 18 of the present invention.
Will be described. This eighteenth embodiment is an embodiment of the invention described in the eighteenth aspect, and in this eighteenth embodiment, before learning, for example, the time distribution of each time zone from 1:00 to 24:00 in the last month is preliminarily set. Calculate the average of the actual water volume. Then, as the input / output data of the neural network,
Weather of the day, maximum temperature of the day, sunshine time of the day, difference between the actual hourly water distribution amount of each time zone of the latest weather day and the hourly average water distribution amount of each time zone, maximum temperature of the most recent weather day, sunshine A total of 29 times are input, and a total of 24 differences between the hourly water distribution in each time zone of the day and the hourly average water distribution in each time zone are output.

For learning of the neural network, the actual value of this input / output data is used as learning data, and the output of the neural network matches the difference between the actual value of the hourly water distribution and the above-mentioned hourly average water distribution. The neural network is learned as described above. In addition, the difference between the forecasted weather, forecasted maximum temperature, forecasted sunshine duration, and the actual amount of hourly water distribution for each time zone of the most recent weather day and the above-mentioned average hourly water distribution will be displayed in the neural network after this learning. By inputting the maximum temperature and the sunshine duration of the same weather day recently, the difference between the hourly water distribution and the hourly average water distribution in each time zone of the predicted day is predicted. Then, based on this predicted value and the hourly average water distribution for each time period, the hourly water distribution for each time period on the predicted day is predicted.

Example 19 As described above, in Examples 11 to 18, the case of predicting the hourly water distribution in each time zone from 1:00 to 24:00 on the predicted day is shown. However, the daily water quantity of the hourly water distribution in each time zone of the predicted day is shown. You may make it estimate the ratio with respect to. The nineteenth embodiment is an embodiment of the invention described in claim 19, and as the input / output data of the neural network, the weather of the day, the maximum temperature of the day, the sunshine time of the day, from 1 o'clock of the most recent weather day. Input the ratio of the actual value of the hourly water distribution for each time zone up to 24:00 to the daily water distribution, the maximum temperature of the most recent weather day, and the sunshine hours of the most recent weather day. The total of the ratio of the hourly water distribution to the daily water distribution for each time zone up to 24 is output.

At the time of learning the neural network, the actual value of this input / output data is used as learning data, and the output of the neural network is adjusted so that it matches the ratio of the actual value of the hourly water distribution on the day to the daily water distribution. The network is learned. Also, at the time of forecasting, forecasted weather on the forecasted day, forecasted maximum temperature on the day, forecasted sunshine hours on the day, ratio of hourly water distribution results to daily distribution on the same day, maximum temperature on the most recent weather day, recent The sunshine duration on the same weather day is input to the neural network after learning, and the ratio of the hourly water distribution to the daily water distribution in each time zone of the prediction day is predicted. Then, based on the predicted value and, for example, the daily water distribution amount obtained by the demand amount prediction method shown in the first embodiment, the hourly water distribution amount in each time zone of the predicted day is predicted.

Example 20. Here, in the above-mentioned Example 19, as the input / output data of the neural network, the weather of the day, the maximum temperature, the sunshine time, the ratio of the actual value of the hourly water distribution in each time zone of the same weather day to the daily water distribution, Although the maximum temperature and sunshine duration on the same weather day were input, the input / output data of the relevant neural network consisted of the ratio of the actual amount of hourly water distribution for each time zone on the same weather day to the daily water distribution, and Instead of the maximum temperature and sunshine hours on the same weather day, the ratio of the hourly water amount actual value to the daily water distribution in each time period from 1:00 to 24:00 of the past T days, and the weather and maximum temperature of the past T days, The sunshine duration may be input, and the same effect as that of the above embodiment is obtained. In this case, at the time of forecasting, forecasted weather, forecasted maximum temperature, forecasted sunshine duration, actual value of ratio of hourly water distribution to daily water distribution in each time zone of the past T days, weather of the last T days, maximum temperature, The sunshine duration is input to the neural network to predict the ratio of the hourly water distribution to the daily water distribution in each time zone of the forecasted day, and the ratio is calculated based on the predicted value and the daily water distribution calculated in Example 2, for example. Predict hourly water distribution for the time zone. In addition, the 20th embodiment is an embodiment of the invention described in claim 20.

Example 21. Next, Example 21 of the present invention
Will be described. The twenty-first embodiment is an embodiment of the invention described in claim 21, and similar to the case of the thirteenth embodiment, before the learning, in advance, as in the case of the thirteenth embodiment, for example, the actual amount of water distributed over the past several months. Based on the weather and the maximum temperature, the data is, for example, three types of weather such as fine, cloudy, and rain, and the maximum temperature is 20 ° C or lower, 20 ° C to 30 ° C, and 30 ° C.
The above three classes are classified into a total of nine classes. Then, prepare nine neural networks corresponding to each of the classes, and in the class including the predicted date, the maximum temperature of the day, as the input / output data to / from the neural network,
Enter 28 sunshine hours on the day, the ratio of the actual amount of water distributed in each time zone of the most recent class day to the daily distribution, the maximum temperature of the most recent class day, and the sunshine hours of the most recent class day. With the output of the total of 24 ratios of the hourly water distribution to the daily water distribution in each time zone, for this input / output relationship, for example, for the data belonging to the same class around the past 8 days, the output of the neural network is the same day. Learning is done so that it matches the ratio of the actual value of hourly water distribution to the daily water distribution.

After the learning is completed, the neural network corresponding to the class including the predicted day of the present day has a daily distribution of the predicted maximum temperature of the predicted day, the predicted sunshine time, and the hourly water distribution amount actual value of the same class day. By inputting the ratio to the amount of water, the maximum temperature of the most recent day in the same class, and the sunshine hours, the ratio of the hourly water distribution to the daily water distribution in each time zone of the forecast day is predicted. Then, based on the predicted value and, for example, the daily water distribution amount obtained by the demand amount prediction method shown in the third embodiment, the hourly water distribution amount in each time zone on the predicted day is predicted.

Example 22. Next, Embodiment 22 of the present invention.
Will be described. The twenty-second embodiment is an embodiment of the invention described in claim 22. Prior to learning, for example, each hourly water distribution result data of the past several months is based on the weather and the maximum temperature, and the weather is fine. , Cloudy, rain, maximum temperature below 10 ℃, 10 ℃ to 13 ℃, 13 ℃
-16 ° C, 16 ° C-19 ° C, 19 ° C-22 ° C, 22 ° C-
25 classes, 25 degrees C-28 degrees C, 28 degrees C-31 degrees C, 31 degrees C or more are classified into a total of 27 classes, and in each class, the ratio of the hourly water distribution results for each time zone to the daily water distribution Take the average and the maximum temperature.

Then, as the input / output data of the neural network, the weather of the class and the average maximum temperature of the class are input, and the average ratio of the hourly water distribution of the class to the daily water distribution is calculated. Using a total of 24 outputs and using the actual values of the above input / output data of all classes as learning data, the output of the neural network matches the average of the ratio of the actual daily distribution of the class to the daily distribution. Learn the neural network as you would. Then, by using the neural network after this learning, by inputting the forecast weather and forecast maximum temperature of the water distribution forecast day, the ratio of the hourly water distribution to the daily water distribution in each time zone of the forecast day can be calculated. Make you predict. Then, based on the predicted value and, for example, the daily water distribution amount obtained by the demand amount prediction method shown in the fifth embodiment, the hourly water distribution amount in each time zone on the predicted day is predicted.

Example 23. Next, Example 23 of the present invention
Will be described. The twenty-third embodiment is an embodiment of the invention described in claim 23, and the input / output data to / from the neural network is the weather difference between the current day and the previous day, the maximum temperature difference between the current day and the previous day, and the current day and the previous day. Difference in sunshine hours, difference between the ratio of the actual amount of hourly water distribution to the daily water distribution in each time zone from 1 o'clock to 24 o'clock between the same weather day and the day before the same weather day, Weather difference from the previous day, maximum temperature difference,
A total of 30 sunshine duration differences are input, and a total of 24 differences in the ratio of the hourly water distribution amount to the daily water distribution amount between 1 o'clock and 24 o'clock on the forecast day and the day before the forecast day are output.

For learning of the neural network, the actual value of this input / output data is used as learning data, and the output of the neural network is the difference between the ratio of the actual value of the hourly water distribution on the current day to the previous water distribution. The neural network is trained so that they match. In addition, the neural network after this learning shows the predicted weather difference between the predicted day and the day before the predicted day, the predicted maximum temperature difference, the predicted sunshine duration difference, the most recent weather day of the predicted day, and the most recent weather day of the predicted day. The difference in the ratio of the actual amount of water distributed in each time zone to the previous day and the amount of water distributed for each day, the difference in weather between the most recent weather day on the forecast day and the day before the most recent weather day on the forecast day, the maximum temperature difference, and the sunshine duration By inputting, the difference in the ratio of the hourly water distribution amount to the daily water distribution amount in each time zone between the prediction day and the day before the prediction day is predicted. Then, the ratio of the hourly water distribution for each time zone to the daily water distribution calculated from the predicted value and, for example, Example 6
Based on the daily water supply volume obtained by the demand forecast method shown in
Forecast The forecast of hourly water distribution in each time zone on the day of the forecast.

Example 24. Next, Embodiment 24 of the present invention.
Will be described. The twenty-fourth embodiment is an embodiment of the invention described in claim 24, and similar to the thirteenth embodiment, before the learning, for example, the hourly water distribution result data of the past several months is recorded in advance for the weather and the maximum temperature. On the basis of, for example, weather is fine, cloudy, rain, and maximum temperature is 20 ℃.
In the following, three classes of 20 ° C. to 30 ° C. and 30 ° C. or more are classified into a total of nine classes, and nine neural networks are prepared for each class. Then, as the input / output data of each of those neural networks, the maximum temperature difference between the current day and the previous same-class day, the sunshine duration difference between the current day and the previous same-class day, the recent same-class day and the previous same-class day The difference in the ratio of the actual amount of hourly water distribution for each time zone from 1 o'clock to 24 o'clock with respect to the day, the maximum temperature difference between the latest same class day and the previous same class day, and the difference in sunshine hours A total of 28 are input, and a total of 24 differences in the ratio of the hourly water distribution amount to the daily water distribution amount for each time zone from 1:00 to 24:00 between the current day and the same class day are output.

The learning of the neural network is performed for the neural network corresponding to the class including the predicted day, and the actual values of the input / output data of the same class for the past several days are used as the learning data. The neural network is trained so that the output matches the difference in the ratio of the hourly water distribution to the daily water distribution of the actual value of the daily water distribution between the same day and the previous class day.
In addition, the neural network after this learning shows the predicted maximum temperature difference between the prediction day and the previous same-class day, the predicted sunshine duration difference, the latest same-class day, and the latest same-class day before the same-class day. By inputting the difference in the ratio of the hourly water distribution amount of each hourly water distribution amount to the daily water distribution amount, the maximum temperature difference between the last same class day and the previous same class day, and the sunshine duration difference, the predicted day of the day The difference in the ratio of the hourly water distribution to the daily water distribution in each time zone is predicted. Then, based on the ratio of the hourly water distribution amount to the daily water distribution amount in each time zone obtained from the predicted value and the daily water distribution amount obtained by the demand amount forecasting method shown in Example 7, for example, each day of the forecast day Predict hourly water distribution in the time zone.

Example 25. Next, Example 25 of the present invention
Will be described. The twenty-fifth embodiment is an embodiment of the invention described in claim 25. In the twenty-fifth embodiment, before learning, the direct fitting method equivalent to that for the daily water distribution amount performed in the eighth embodiment is used. the weather,
The actual data of the ratio of the hourly water distribution to the daily water distribution is classified according to the maximum temperature and the day of the week, and it is 1:00 as the average of the ratio of the hourly water distribution to the daily water distribution in the same class, for example, the past three days. Predict the ratio of hourly water distribution to daily water distribution for each hour from 12:00 to 24:00. Then, as the input / output data of the neural network, the daily weather, the maximum temperature of the day, the sunshine time of the day, and the daily distribution of the hourly water distribution in each time zone from 1 o'clock to 24 o'clock according to the prediction method of the latest day of the same day Prediction error of the ratio to the amount of water, maximum temperature of the most recent weather day, total of 29 sunshine hours of the most recent weather day are input, and the day of the hourly water distribution in each time zone from 1 o'clock to 24 o'clock according to the prediction method of the day A total of 24 prediction errors of the ratio to the amount of water distributed are output.

For learning of the neural network, the actual value of this input / output data is used as learning data, and the output of the neural network is the prediction error of the ratio of the hourly water distribution to the daily water distribution by the prediction method of the day. The neural network is learned so as to match the actual value of. In addition, the forecasted weather on the forecast day, the forecasted maximum temperature, the forecasted sunshine time, and the ratio of the hourly water distribution to the daily water distribution in each time zone according to the above-mentioned prediction method of the same weather day are displayed in the neural network after this learning. By inputting the prediction error, the maximum temperature of the same weather day, and the sunshine time,
A prediction error of the ratio of the hourly water distribution to the daily water distribution in each time zone of the prediction day is predicted by the prediction method. Then, based on the ratio of the hourly water distribution for each time zone to the daily water distribution calculated from the predicted value and the daily water distribution calculated by the demand prediction method shown in Example 8, for example, each day of the prediction day Predict hourly water distribution in the time zone.

Example 26. Next, Embodiment 26 of the present invention.
Will be described. This twenty-sixth embodiment is an embodiment of the invention described in claim 26. In this twenty-sixth embodiment, the time distribution of each time period from 1 o'clock to 24 o'clock in the last month in advance is performed before learning. The average of the ratio of the actual amount of water supply to the daily distribution is calculated. And, as the input and output data of the neural network, the weather of the day, the maximum temperature of the day,
The sunshine hours on the day, the difference between the ratio of the actual value of the hourly water distribution for each time zone to the daily water quantity and the average ratio of the hourly water quantity to the daily water quantity for each time zone, and the most recent day of the same weather day Enter the maximum temperature and the total of 29 sunshine hours on the most recent weather day, and calculate the ratio of the hourly water distribution in each time zone to the daily water quantity and the average ratio of the hourly water quantity to the daily water quantity in each time zone. A total of 24 differences are output.

For learning of the neural network, the actual values of this input / output data are used as learning data, and the output of the neural network is the ratio of the actual value of the hourly water distribution to the daily water distribution of the day and the time distribution of each time zone. The neural network is trained so as to match the difference between the above-mentioned hourly water distribution amount and the average ratio of the daily water distribution amount. In addition, the forecasted weather, forecasted maximum temperature, forecasted sunshine time, and the ratio of the actual amount of hourly water distribution to the daily water distribution for each time zone on the same weather day are shown in the neural network after this learning. The ratio of the hourly water distribution to the daily water distribution in each time zone of the forecasted day by entering the difference between the above-mentioned hourly water distribution and the average ratio to the daily water distribution, the maximum temperature of the most recent weather day, and the sunshine duration. And predict the difference between the above-mentioned hourly water distribution amount and the average ratio of the daily water distribution amount in each time zone. Then, based on the ratio of the hourly water distribution for each time zone to the daily water distribution calculated from the predicted value and the daily water distribution calculated by the demand prediction method shown in Example 10, for example, each day of the predicted day Predict hourly water distribution in the time zone.

Example 27. Next, Embodiment 27 of the present invention.
Will be described with reference to FIG. The twenty-seventh embodiment is an embodiment of the invention set forth in claim 27, in which the learning parameter is searched for a learning parameter having the smallest average prediction error while sequentially changing the learning parameter, and the learning parameter is used to search the neural network. Learning and prediction of the amount of water distribution, etc. are performed using the neural network after learning, and the flow of the operation is shown in FIG.
0 is shown in the flowchart.

In FIG. 20, after the learning end condition is initialized to 1.5% in step ST1, step ST2
Then, it is monitored whether or not this learning processing condition exceeds 2.4%. If the learning condition is less than 2.4%, step ST3
Initialize the number of learning samples to 11 in step ST
While monitoring whether or not the learning sample number exceeds 20 in 4, the daily water distribution amount described in the first embodiment is predicted in step ST5, and the learning sample number is incremented in step ST6. When it is detected in step ST4 that the number of learning samples exceeds 20, step ST7
Then, the learning end condition is increased by 0.1% and the process returns to step ST2. Therefore, the learning condition is 2.4% in this step ST2.
The above process is repeated until it is detected that When it is detected in step ST2 that the learning condition exceeds 2.4%, the number of learning samples for which the average prediction error is minimized and the learning end condition are displayed in step ST8, and then the process ends. In this way, the average prediction error of the learning sample which is the learning end condition is sequentially increased from 1.5% to 2.4% at every 0.1% while increasing the number of learning samples from 11 to 20 one by one. The prediction in Example 1 is performed by changing it.

From this result, the number of learning samples and the learning end condition that minimize the average prediction error are obtained, and the prediction described in the first embodiment is performed using the obtained learning sample number and learning end condition. Incidentally, FIG. 21 shows the relationship between the learning sample number and the learning end condition, and the average prediction error. In this graph, the learning sample number and the learning end condition that minimize the average prediction error are obtained.

Example 28. Next, Embodiment 28 of the present invention.
Will be described. This embodiment 28 is an embodiment of the invention described in claim 28, and in this embodiment 28, the actual value of the input / output data excluding the days when the actual value of the amount of water distribution greatly differs from the actual value of the amount of water distribution before and after is shown. The neural network is learned as learning data, and the daily water distribution and the hourly water distribution are predicted using the neural network after the learning is completed. That is, the neural network is learned by excluding the days on which the actual distribution values differ significantly from the previous day and the next day from the learning sample. Specifically, data that satisfies the following equation (13) Only the learning sample is adopted.

Q ₁ + Q ₂ <α + (W ₁ + W ₂ ) × β (13) Q ₁ : Absolute value of difference of water distribution from the previous day Q ₂ : Difference of water distribution from the following day Absolute value W ₁ : Absolute value of difference of weather coefficient from the previous day W ₂ : Absolute value of difference of weather coefficient from the following day α, β: Constants

[0146]

As described above, according to the invention of claim 1, the weather, the maximum temperature, the sunshine time of the day, the actual daily demand amount of the most recent weather day, the maximum temperature, and the sunshine time are input. The daily demand is output and the input / output relationship is learned by the neural network to predict the daily demand. Therefore, the relationship between the factors affecting the demand and the daily demand is Non-linear approximation can be performed using a circuit network, which has the effect of predicting daily demand with high prediction accuracy.

According to the second aspect of the present invention, the weather, maximum temperature, sunshine duration on the day, actual daily demand amount for the past predetermined number of days, weather, maximum temperature, sunshine duration are input, and the daily demand on the day is entered. The output is used as the output, and the input / output relations are learned by the neural network to predict the daily demand. Therefore, the relationship between the factors affecting the demand and the daily demand is calculated by the neural network. It is possible to make a non-linear approximation by using it, and it is possible to predict daily demand with high prediction accuracy.

Further, according to the invention of claim 3, before learning, the past daily demand data is divided into classes by weather and maximum temperature, and a neural network is prepared for each class. , The maximum temperature of the day, sunshine hours, actual daily demand of the same class in the class including the forecast day,
The maximum temperature and sunshine hours are input, the daily demand is output, and the input / output relationship of the days that belong to the class that includes the forecast date is learned by the neural network corresponding to the class that includes the forecast date. Since it is configured to predict the daily demand, it is possible to nonlinearly approximate the relationship between the factors affecting the demand and the daily demand using the neural network, and predict the daily demand with high prediction accuracy. There is an effect that can be.

According to the invention described in claim 4, the daily demand result data is standardized in advance by a linear approximation before learning, and the current day's weather, maximum temperature, sunshine duration, and recent similar weather day are standardized. It is configured to predict the standardized daily demand amount by inputting the actual daily demand amount, the maximum temperature, and the sunshine duration, and using the standardized daily demand amount of the day as the output, and learning the input / output relationship to the neural network. Therefore, it becomes possible to nonlinearly approximate the relationship between the factors affecting the demand amount and the standardized daily demand amount by using the neural network, and the daily demand amount with higher prediction accuracy than the predicted standardized daily demand amount. There is an effect that can be predicted.

According to the invention of claim 5, before learning, the past daily demand data is divided into classes by weather and maximum temperature, and the average daily demand and maximum temperature are averaged in each class. Then, the weather of the class and the average maximum temperature are input, the average daily demand of the class is output, and the neural network is made to learn the above-mentioned input / output relationship of all classes to predict the daily demand. Therefore, with noise removed as much as possible, the relationship between factors affecting demand and daily demand can be nonlinearly approximated using the neural network, and daily demand can be predicted with high prediction accuracy. There is an effect that can be.

According to the invention described in claim 6, the difference in weather between the current day and the day before, the difference in maximum temperature, the difference in sunshine duration, the difference in actual demand amount between the most recent same day and the day before the most recent same day, the difference in weather ,
Since the maximum temperature difference and the sunshine duration difference are input, and the difference in daily demand between the current day and the previous day is output, the input / output relationship is learned by the neural network to predict the difference in daily demand between the current day and the previous day. The relationship between the factors affecting the demand and the daily demand difference can be nonlinearly approximated by using the neural network, and the daily demand amount can be predicted with higher prediction accuracy than the predicted daily demand difference. There is an effect that can be.

According to the invention described in claim 7, before the learning, the past daily demand result data is divided into classes by weather and maximum temperature, and a neural network is prepared for each class. , In the class that includes the forecast date, the maximum temperature difference between the current day and the previous same-class day, the sunshine duration difference, the actual difference in daily demand amount between the latest same-class day and the previous same-class day of the latest same-class day, the maximum temperature difference, Input the sunshine time difference and output the daily demand difference between the current day and the same class day as the output, and train the neural network corresponding to the class including the prediction date to learn the above-mentioned input / output relation of the day belonging to the class including the prediction date. Since it is configured to predict the daily demand difference between the current day and the previous same-class day, the neural network can be used to nonlinearly approximate the relationship between the factors affecting the demand and the daily demand difference. It becomes possible, and from the predicted daily demand difference There is an effect that it is possible to predict the day demand with high prediction accuracy.

Further, according to the invention described in claim 8, the daily demand amount is predicted in advance by the direct fitting method before the learning, and the weather of the day, the maximum temperature, the sunshine time, and the recent weather day are predicted. Input the daily demand forecast error, the maximum temperature, and the sunshine time, and use the daily demand forecast error of the current day's forecast as an output, and learn the input / output relationship to the neural network to forecast the daily demand by the forecast method. Since it is configured to predict the error, it becomes possible to perform a non-linear approximation of the relationship between the factors that affect the demand amount and the prediction error of the daily demand amount by the prediction method using the neural network. There is an effect that the daily demand amount can be predicted with higher prediction accuracy than the predicted error.

According to the ninth aspect of the invention, the daily demand is predicted in advance by linear approximation before learning,
Input the weather, maximum temperature, sunshine time on the day, forecast error of daily demand amount by the above-mentioned forecast of the latest weather day, maximum temperature, sunshine time, and output the error of forecast daily demand amount by the forecast on the day, and its input / output relationship Since it is configured to predict the prediction error of the daily demand by the prediction method by learning the neural network, the relationship between the factors that affect the demand and the prediction error of the daily demand by the prediction method will be described. , It becomes possible to perform non-linear approximation using the neural network, and there is an effect that the daily demand amount can be predicted with higher prediction accuracy than the predicted prediction error.

According to the invention described in claim 10,
Before learning, the average of the daily demand actual value for a predetermined period is calculated in advance, and the weather on that day, the maximum temperature, the sunshine time, the difference between the actual daily demand amount on the same weather day and the average daily demand, the maximum temperature,
Input the sunshine duration, output the difference between the daily demand amount and the average daily demand amount, and train the neural network for the input / output relationship to predict the difference between the current day demand amount and the average daily demand amount. Since it has been configured, it becomes possible to nonlinearly approximate the relationship between the factors affecting the demand amount and the difference between the daily demand amount and the average daily demand amount by using the neural network, and the predicted average daily demand There is an effect that the daily demand amount can be predicted with high prediction accuracy from the difference with the amount.

According to the invention described in claim 11,
Input the weather, maximum temperature, sunshine time, actual hourly demand amount for each time zone of the same weather day, maximum temperature, sunshine hour, and output hourly demand amount for each day on that day. Since the neural network is configured to learn and predict the time demand amount in each time zone, the relationship between the factors affecting the demand amount and each time demand amount is nonlinearly approximated using the neural network. This makes it possible to predict the time demand amount in each time zone with high prediction accuracy.

Further, according to the invention of claim 12,
Input the weather, maximum temperature, sunshine hours on the day, actual demand for each time zone during the past predetermined number of days, weather, maximum temperature, sunshine hours, and output the hourly demand for each time zone on that day. Since the neural network learns the relationship and predicts the time demand in each time zone, the relationship between the factors affecting the demand and each time demand can be calculated using the neural network. It becomes possible to make an approximation, and there is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy.

According to the invention described in claim 13,
Before learning, classify the past hourly demand data by weather and maximum temperature, prepare a neural network for each class, and in the class that includes the forecast date,
Input the maximum temperature, sunshine hours, actual hourly demand amount for each time zone of the same class day, maximum temperature, sunshine hours, and output the hourly demand amount for each time zone on the day, and classify into the class that includes the forecast date. The input / output relationship of the day to which it belongs is configured to predict the hourly demand for each time zone by learning each neural network corresponding to the class that includes the forecasted day. The relationship with the demand amount can be nonlinearly approximated by using the neural network, and the time demand amount in each time zone can be predicted with high prediction accuracy with high prediction accuracy.

According to the invention of claim 14,
Prior to learning, the past hourly demand data is divided into classes by weather and maximum temperature, and the average of the hourly demand data and maximum temperature for each time zone is calculated for each class, and the weather and average maximum temperature of the class are calculated. , And the time average demand for each time zone of the class is output, and the neural network is made to learn the above-mentioned input / output relations of all classes to predict the time demand of each time zone. It is possible to nonlinearly approximate the relationship between factors affecting demand volume and each hourly demand volume by using the neural network, with high removal accuracy. There are predictable effects.

Further, according to the invention of claim 15,
Enter the weather difference, maximum temperature difference, sunshine duration difference between the day and the previous day, the actual demand difference, the weather difference, the maximum temperature difference, the sunshine duration difference for each time zone on the day before the same weather day and the day before the same weather day. As the output is the difference in the time demand amount for each time zone on the previous day, the input / output relationship is learned by the neural network to predict the time demand amount difference for each time period. And the relationship between the hourly demand difference on the current day and the previous day can be nonlinearly approximated using the neural network.
There is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy from the predicted difference in each time demand amount.

According to the invention described in claim 16,
Before learning, classify the past hourly demand data by weather and maximum temperature, prepare a neural network for each class, and in the class that includes the forecast date,
Enter the maximum temperature difference between the current day and the previous same class day, the sunshine duration difference, the actual difference in hourly demand amount for each time zone between the latest same class day and the previous previous class day, the maximum temperature difference, and the sunshine duration difference,
Output the time demand difference between each time zone of the current day and the same class day before, and train the neural network corresponding to the class including the prediction date to learn the above-mentioned input / output relationship of the day belonging to the class including the prediction date. Since it is configured to predict the time demand difference in each time zone, the relationship between the factors that affect the demand amount and the time demand difference between the current day and the previous class day is calculated using the neural network. Non-linear approximation is possible, and there is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy from the predicted time demand amount difference.

According to the invention of claim 17,
Similar to the direct fitting method in advance before learning, predict the time demand of each time zone, the weather on the day, the maximum temperature, the sunshine time, the time demand forecast error of each time zone by the forecast of the latest weather day, The maximum temperature and sunshine hours are input, and the time demand forecast error of each time zone is output based on the forecast of the current day. The neural network learns the input / output relationship and forecasts the time demand forecast error of each time zone. Since it is configured so that it is possible to nonlinearly approximate the relationship between the factors that affect the demand amount and the prediction error of each time demand amount by the prediction method using the neural network. From the prediction error, there is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy.

According to the invention as set forth in claim 18,
Before learning, the average of the hourly demand amount actual value of each time period in a predetermined period is calculated in advance, and the actual day's weather, maximum temperature, sunshine time, and recent hourly demand amount actual time average amount and hourly average demand Input the difference from the amount, maximum temperature, and sunshine hours, and output the difference between the hourly demand amount and the hourly average demand amount of each time zone of the day, and let the neural network learn the input / output relationship, and Since it is configured to predict the difference between the hourly demand amount and the hourly average demand amount, the factors that affect the demand amount and the relationship between the hourly demand amount of the day and the difference between the hourly average demand amounts are Non-linear approximation can be performed using a neural network, and there is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy from the difference between the predicted time average demand amounts.

According to the invention of claim 19,
Weather of the day, maximum temperature, sunshine hours, ratio of hourly demand to daily demand in each time zone of the latest weather day, maximum temperature,
Input the sunshine time, output the ratio of the hourly demand to the daily demand in each time zone of the day, train the neural network to learn the input / output relationship, and calculate the ratio of the hourly demand to the daily demand in each time zone. Since it is configured to predict, the relationship between the factors that affect demand and the ratio of each hourly demand to the daily demand can be nonlinearly approximated by using the neural network. From the ratio to daily demand,
There is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy.

Further, according to the invention of claim 20,
Enter the weather, maximum temperature, sunshine hours, ratio of hourly demand for each time period to the daily demand for the past predetermined number of days, weather, maximum temperature, sunshine time, day of hourly demand for each day of the day The output is a ratio to the demand amount, and the input / output relationship is learned by the neural network to predict the ratio of the time demand amount to the daily demand amount in each time zone. , The relationship between the ratio of each hourly demand to the daily demand can be nonlinearly approximated using the neural network, and the hourly demand in each time zone is higher than the predicted daily demand. There is an effect that can be predicted with the prediction accuracy.

Further, according to the invention of claim 21,
Before learning, classify the past hourly demand data by weather and maximum temperature, prepare a neural network for each class, and in the class that includes the forecast date,
Enter the maximum temperature, sunshine hours, the ratio of hourly demand to the daily demand in each time zone of the same class day, enter the maximum temperature and sunshine hours, and the ratio of the hourly demand to the daily demand in each time zone of the day. Is output, and the neural network corresponding to the class including the prediction day is made to learn the above-mentioned input / output relation of the day belonging to the class including the prediction day, and the ratio of the time demand amount to the daily demand amount in each time zone is predicted. With this configuration, it becomes possible to use the neural network to nonlinearly approximate the relationship between the factors that affect the demand volume and the ratio of each hourly demand volume to the daily demand volume. There is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy based on the ratio to the amount.

Further, according to the invention of claim 22,
Before learning, classify the past hourly demand actual data into classes by weather and maximum temperature, and calculate the ratio of hourly demand to daily demand in each time zone and the average of maximum temperature in each class. The weather and average maximum temperature are input, and the average of the ratio of the hourly demand to the daily demand in each time zone of the class is output, and the neural network learns the input / output relationship of all classes, and the time of each time zone is learned. Since it is configured to predict the average of the demand amount to the daily demand amount, the relationship between the factors that affect the demand amount and the ratio of the hourly demand amount to the daily demand amount is calculated with noise removed as much as possible. , It becomes possible to perform non-linear approximation using the neural network, and it is possible to predict the time demand amount in each time zone with high prediction accuracy from the ratio to the predicted daily demand amount.

Further, according to the invention of claim 23,
Weather difference between the day and the day before, maximum temperature difference, sunshine difference, difference between the ratio of the actual demand amount to the daily demand amount, the difference in weather, the maximum temperature difference, the difference in lighting time The output is the difference in the ratio of the hourly demand to the daily demand for the current day and the previous day, and the neural network learns the input / output relationship to determine the ratio of the daily demand to the hourly demand for the current day and the previous day. Since it is configured to predict the difference, the relationship between the factor affecting the demand amount and the difference in the ratio of the hourly demand amount of the current day to the daily demand amount of the previous day is nonlinearly approximated using the neural network. This makes it possible to predict the time demand amount in each time zone with high prediction accuracy from the difference in the ratio to the predicted daily demand amount.

According to the invention of claim 24,
Before learning, classify the past hourly demand data by weather and maximum temperature, prepare a neural network for each class, and in the class that includes the forecast date,
Maximum temperature difference between the current day and the previous same class day, sunshine duration difference, difference in the ratio of hourly demand to daily demand in each time zone between the latest same class day and the previous previous class day, the maximum temperature difference, Input the difference in sunshine hours, and output the difference in the ratio of the hourly demand amount to the daily demand amount for each time zone of the current day and the same class day before,
Ratio of the hourly demand to the daily demand in each time zone of the current day and the previous same-class day by learning the above-mentioned input / output relationship of the day belonging to the class including the forecasted day by the neural network corresponding to the class including the forecasted day Since it is configured to predict the difference between the demand and the demand, the relationship between the factors affecting the demand and the difference in the ratio of the hourly demand to the daily demand on the same day and the previous same class is calculated using the neural network. It is possible to make a non-linear approximation, and there is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy from the difference in the ratio to the predicted daily demand amount.

According to the invention of claim 25,
Prior to learning, predict the ratio of the time demand amount to the daily demand amount in each time zone in advance as in the direct fitting method in advance, and learn the weather of the day, the maximum temperature, the sunshine time, the most recent day of the same weather Input and output the prediction error of the ratio of hourly demand to daily demand, the maximum temperature, and sunshine hours, and the output of the prediction error of the ratio of hourly demand to daily demand in each time zone according to the forecast of the current day. Since the neural network is trained to predict the prediction error of the ratio of each hourly demand to the daily demand by the above prediction,
The relationship between the factor affecting the demand amount and the prediction error of the ratio of each time demand amount to the daily demand amount by the prediction can be nonlinearly approximated by using the neural network, and the predicted daily demand can be obtained. From the prediction error of the ratio to the quantity,
There is an effect that the time demand amount in each time zone can be predicted with high prediction accuracy.

Further, according to the invention of claim 26,
Before learning, find the average of the ratio of the actual demand value for each time period of the predetermined period to the daily demand amount in advance, and find the weather on that day, the maximum temperature, the sunshine time, and the time demand for each time period on the most recent weather day. The difference between the ratio to the daily demand amount and the average value, the maximum temperature, the sunshine hours are input, and the difference between the average value and the ratio to the daily demand amount of the time demand amount in each time zone of the day is input, Since the input / output relationship is learned by the neural network and the difference between the average value and the ratio of the time demand amount to the daily demand amount in each time zone is configured to be predicted, the factors that affect the demand amount, The relationship between the ratio of the hourly demand amount to the daily demand amount by the prediction and the difference between the average values can be nonlinearly approximated by using the neural network, and the ratio to the predicted daily demand amount and the average value can be obtained. Depending on the difference from the value, There is an effect that it is possible to predict time demand with high prediction accuracy.

Further, according to the invention of claim 27,
While sequentially changing the learning parameters, learning and prediction are performed by the neural network, the value of the learning parameter that minimizes the average prediction error is obtained, and the daily demand forecast or time demand forecast is performed with the learning parameter value. Because I configured
There is an effect that the prediction accuracy of the daily demand amount and the hourly demand amount can be further improved.

Further, according to the invention of claim 28,
Since the neural network is configured to be trained by excluding the days when the actual demand value is significantly different from the actual demand values before and after the learning data, noise included in the learning data can be removed. There is an effect that it is possible to further improve the prediction accuracy of the quantity and time demand.

[Brief description of drawings]

FIG. 1 is a conceptual diagram showing a system to which a demand quantity forecasting method according to a first embodiment of the present invention is applied.

FIG. 2 is a conceptual diagram showing a hierarchical neural network in the above embodiment.

FIG. 3 is a conceptual diagram showing a recurrent neural network in the above embodiment.

FIG. 4 is a conceptual diagram showing chaotic neurons of the chaotic neural network in the above embodiment.

FIG. 5 is a conceptual diagram for explaining a demand quantity forecasting method according to a third embodiment of the present invention.

FIG. 6 is an explanatory diagram showing a prediction result by a conventional direct fitting method.

FIG. 7 is an explanatory diagram showing a prediction error by a conventional direct fitting method.

FIG. 8 is an explanatory diagram showing a prediction result according to the method of the first embodiment.

FIG. 9 is an explanatory diagram showing a prediction error according to the method of the first embodiment.

FIG. 10 is an explanatory diagram showing a prediction result according to the method of the third embodiment.

FIG. 11 is an explanatory diagram showing a prediction error according to the method of the third embodiment.

FIG. 12 is an explanatory diagram collectively showing prediction errors by the above methods.

FIG. 13 is an explanatory diagram showing the cumulative number of days of the prediction error of each method.

FIG. 14 is an explanatory diagram showing the average prediction error of each method in a table format.

FIG. 15 is a conceptual diagram for explaining a demand quantity prediction method according to a fourth embodiment of the present invention.

FIG. 16 is a conceptual diagram showing input / output data of a neural network according to the sixth embodiment of the present invention.

FIG. 17 is a conceptual diagram showing input / output data of the neural network according to the seventh embodiment of the present invention.

FIG. 18 is a conceptual diagram showing input / output data of the neural network according to the eighth embodiment of the present invention.

FIG. 19 is a conceptual diagram showing input / output data of a neural network according to the tenth embodiment of the present invention.

FIG. 20 is a flow chart showing an operation flow of a demand quantity prediction method according to a twenty-seventh embodiment of the present invention.

FIG. 21 is an explanatory diagram showing the relationship between the learning end condition, the learning sample number, and the average prediction error in the above embodiment.

[Explanation of symbols]

1 water flow rate detector, 2 day demand forecasting device, 3 day demand forecasting value, 4 hour water demand forecasting device, 5 water distribution control device, 6 demand amount actual value file, 7 meteorological related actual data file, 9 weather forecasting data Files, 10 normalizers, 11 neural networks.

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Makoto Tsukiyama 8-1-1 Tsukaguchihonmachi, Amagasaki City Mitsubishi Electric Corporation Industrial Systems Research Center

Claims (28)

[Claims] 1. A demand quantity forecasting method for forecasting a demand quantity which fluctuates depending on weather conditions by using a neural network, wherein as input data of input / output data inputted / outputted to / from the neural network, weather and maximum temperature of the day. , Sunshine hours, and the daily demand results of the same weather day, the maximum temperature, and the sunshine hours are used as the output data, and the daily demand amount is used as the output data. The actual value of the input / output data within a certain period is used as the learning data. , So that the output of the neural network and the actual daily demand value match,
The neural network performs learning, the neural network after the learning is completed, forecasted weather for forecasting the demand amount, forecasted maximum temperature, forecasted sunshine duration, and daily demand results for the most recent weather day, A demand quantity forecasting method, which comprises predicting a daily demand quantity on the forecast day by inputting a maximum temperature and a sunshine duration. 2. A demand quantity forecasting method for forecasting a demand quantity which fluctuates depending on weather conditions using a neural network, wherein the input data of the input / output data to / from the neural network is the weather and maximum temperature of the day. , Sunshine hours, and daily demand results for a certain number of days in the past, weather, maximum temperature, sunshine hours, and using the daily demand amount as output data, learn the actual values of the input / output data within a certain period As the data, the neural network is made to perform learning so that the output of the neural network and the actual daily demand value match, and the neural network after the learning is completed has a forecast date for demand forecasting. Predicted weather, predicted maximum temperature, predicted sunshine duration, and daily demand actual results for the predetermined number of days, weather, maximum temperature, and sunshine duration are input to predict the daily demand for the predicted day. Quantity prediction method. 3. A demand forecasting method for forecasting demand varying according to weather conditions using a neural network, wherein the neural network is classified into past daily demand actual data based on weather and maximum temperature. Prepared for each of the classes, in the class that includes the forecast day for which the demand forecast is performed, as the input data of the input and output data input to and output from the neural network, the maximum temperature of the day, the sunshine duration, and Recent daily demand for the same class, weather, maximum temperature,
Using the sunshine hours as the output data and the daily demand for the day,
Using the actual value of the input / output data for the same number of days of the same class in the past as learning data, the neural network performs learning so that the output of the neural network and the actual daily demand value match. After the learning, enter the predicted maximum temperature, predicted sunshine time, and actual daily demand for the same class date, maximum temperature, and sunshine time into the neural network to calculate the daily demand for the predicted day. A demand forecasting method characterized by forecasting. 4. A demand forecasting method for forecasting a demand varying according to weather conditions using a neural network, wherein daily demand actual data is standardized by linear approximation before learning of the neural network. , As the input data of the input / output data input / output to / from the neural network, the weather of the day, the maximum temperature, the sunshine time, and the standardized daily demand record of the most recent weather day, the maximum temperature, and the sunshine time are output. Use the standardized daily demand for the day as data, and use the actual value of the input / output data for a certain period as learning data so that the output of the neural network matches the standardized daily demand actual value. In addition, the neural network is trained, and the neural network after the learning is completed, the forecasted weather for forecasting the demand amount, the forecasted maximum temperature, the forecasted sunshine duration, and the standard for the most recent weather day Inputting the actualized daily demand amount, maximum temperature, and sunshine duration, predicting the standardized daily demand amount on the forecast day, and predicting the daily demand amount by the linear approximation from the predicted value. Demand forecast method. 5. A demand forecasting method for forecasting a demand varying according to weather conditions by using a neural network, wherein the past daily demand actual data is set in advance as weather and maximum temperature before learning of the neural network. Classifying based on the class, the average of the daily demand results and the maximum temperature is obtained for each class, the weather of the class, the average maximum temperature, as the input data of the input and output data input and output to the neural network, The average daily demand of the class is used as output data, and the actual values of the input / output data of all classes are used as learning data so that the output of the neural network and the average daily demand of the class match. , The neural network is made to perform learning, and after the learning is completed, the forecasted weather and forecasted maximum temperature at which the demand forecast is made are input to the neural network to forecast the daily demand on the forecasted date. Demand forecasting method characterized by measuring. 6. A demand quantity forecasting method for forecasting a demand quantity which fluctuates depending on weather conditions by using a neural network, wherein the difference in weather between the current day and the day before is used as input data of input / output data to / from the neural network. , Maximum temperature difference, sunshine hours difference,
And recently the same day and the previous day's day demand actual difference, weather difference, maximum temperature difference, sunshine duration difference, using the daily demand difference of the current day and the previous day as output data, the input and output data of a certain period Using the actual value as learning data, the neural network is made to perform learning so that the output of the neural network and the daily demand actual value difference of the current day and the previous day match, and the neural network after learning is completed. Forecasting demand, forecasted weather difference on the day before and the day before, forecasted maximum temperature difference, forecasted sunshine duration difference, and forecasted recent day on the same weather day and the previous day's actual demand difference, weather difference, maximum temperature difference A demand amount forecasting method, characterized in that a difference in sunshine duration is input to predict a difference in daily demand amount between the forecasted day and the preceding day. 7. A demand forecasting method for forecasting a demand varying according to weather conditions using a neural network, wherein the neural network is classified into past daily demand actual data based on weather and maximum temperature. Prepared for each class, and in the class that includes the forecast day when the demand forecast is performed, as the input data of the input and output data that is input to and output from the neural network, the maximum temperature on the same day and the previous same day Difference, difference in sunshine hours, and the difference in actual demand amount between the latest same class day and the previous same class day, the maximum temperature difference,
The difference in sunshine duration is used as the output data for the daily demand difference between the current day and the previous same class day, and the actual value of the input / output data for a certain number of days in the same class day in the past is used as learning data, and the output of the neural network and the current day So that the difference in the actual daily demand value of the same class day is matched, the neural network is made to perform learning, and the neural network after the learning is completed Enter the forecasted maximum temperature difference, forecasted sunshine duration difference on the day before the same class day, and the daily demand actual difference, maximum temperature difference, and sunshine duration difference on the previous same class day between the latest same class day and the latest same class day, A demand forecasting method, which comprises forecasting a daily demand difference between the current day of the forecast date and the same class day before the forecast date. 8. A demand forecasting method for forecasting a demand varying according to weather conditions using a neural network, in advance of learning of the neural network, in advance, weather, maximum temperature, day of the week, daily demand actual data. , The daily demand amount is predicted by a direct fitting method that takes the average of the past predetermined number of days belonging to the same class as the predicted value, and is used as the input data of the input / output data input / output to the neural network. ,
Weather, maximum temperature, sunshine duration on the day, and daily demand forecast error due to the above-mentioned forecast of the latest weather day, maximum temperature, sunshine duration, using the daily demand forecast error due to the above-mentioned forecast as output data, As the learning value of the actual value of the input / output data in the period of, the neural network is allowed to perform learning so that the output of the neural network and the actual value of the daily demand forecast error due to the prediction match. After the learning, in the neural network, the forecast weather for forecasting the demand forecast,
Inputting the predicted maximum temperature, the predicted sunshine duration, and the daily demand forecast error due to the forecast of the latest weather day, the maximum temperature, and the sunshine time, and predicting the forecast error due to the forecast of the daily demand amount on the forecast day. Demand forecasting method. 9. A demand forecasting method for forecasting a demand varying with weather conditions using a neural network, wherein the daily demand is forecasted in advance by linear approximation before learning of the neural network. As the input data of the input / output data input / output to the neural network, the weather data, the maximum temperature, and the sunshine time of the day, and the daily demand forecast error, the maximum temperature, and the sunshine time based on the forecast of the most recent weather day are output data. As the daily demand forecast error by the forecast as the current day, the actual value of the input and output data of a certain fixed period as learning data, the output of the neural network and the actual value of the daily demand forecast error by the forecast and So as to match, the neural network is made to perform learning, and the neural network after the learning is completed is predicted to have a forecasted weather, forecasted maximum temperature, forecasted sunshine duration, and forecasted day. A demand forecasting method, comprising inputting a daily demand forecast error, a maximum temperature, and a sunshine duration based on the forecast on the same weather day, and predicting a forecast error based on the forecast of the daily demand on the forecast date. 10. A demand quantity forecasting method for forecasting a demand quantity which fluctuates according to weather conditions by using a neural network, wherein an average of actual daily demand values for a predetermined period is obtained in advance before learning of the neural network. Every day, as the input data of the input and output data input to and output from the neural network, the weather of the day,
The maximum temperature, sunshine duration, and the difference between the actual daily demand on the same weather day and the calculated average daily demand, the maximum temperature, and the sunshine duration are the output data, and the difference between the daily demand and the average daily demand. Using the actual value of the input / output data for a certain period as learning data, the output of the neural network and the difference between the actual value of the daily demand and the average daily demand are matched so that The neural network is made to perform learning, and the neural network after the learning is completed, the forecasted weather for forecasting the demand amount, the forecasted maximum temperature, the forecasted sunshine duration, and the daily demand record of the most recent weather day and the above A demand forecasting method, comprising inputting a difference in average daily demand, a maximum temperature, and a sunshine duration, and predicting a difference between the demand on the forecast day and the average daily demand. 11. A demand forecasting method for forecasting a demand varying with weather conditions by using a neural network, wherein the input data of input / output data to / from the neural network is the weather and maximum temperature of the day. , Sunshine hours, and the actual hourly demand amount for each time zone of the most recent weather day, maximum temperature, and sunshine hours are used as output data, the hourly demand amount for each time zone of the day is used, and the input / output data for a certain period As the learning data, the neural network is made to perform learning so that the output of the neural network and the actual value of the demanded time coincide with each other, and the neural network after the learning is predicted to have a demand amount. Enter the forecast weather, forecast maximum temperature, forecast sunshine duration, and hourly demand for the most recent weather day, maximum temperature, and sunshine duration to calculate the hourly demand in each time zone of the forecast date. To predict And a demand forecasting method. 12. A demand forecasting method for forecasting a demand varying with weather conditions using a neural network, wherein the input data of input / output data to / from the neural network is the weather and maximum temperature of the day. , Sunshine hours, and hourly demand data for each time period during the past specified number of days, weather, maximum temperature, and sunshine hours are used as output data, and the hourly demand amount for each time zone of the day is used as the input for a certain period. Using the actual value of the output data as the learning data, the neural network is made to perform learning so that the output of the neural network and the actual value of the time demand amount match, and the neural network after the learning is completed Enter the forecast weather, forecast maximum temperature, forecast sunshine duration, and hourly demand data for each of the specified number of days, weather, maximum temperature, and sunshine duration for each forecast day on which the forecast date Time demand A demand forecasting method characterized by forecasting a quantity. 13. A demand forecasting method for forecasting a demand varying according to weather conditions by using a neural network, wherein the neural network is used for past time demand actual data based on the weather and maximum temperature of the day. Prepared for each of the divided classes, in the class including the forecast day for which the demand forecast is performed, as the input data of the input and output data to and from the neural network, the maximum temperature of the day,
The sunshine hours and the actual hourly demand amount for each time zone of the same class day, maximum temperature, and sunshine hour are used as output data, and the hourly demand quantity of each time zone of the current day is used as the output data, before the specified number of days before the same class day. The actual value of the input output data as learning data, so that the output of the neural network and the actual value of the time demand amount match, the neural network is made to perform learning, and the neural network after the learning is completed, Enter the forecast maximum temperature, forecast sunshine duration, and hourly demand amount for each recent class day, maximum temperature, and sunshine duration to forecast the maximum demand temperature in each time zone on the forecast date. A demand forecasting method characterized by forecasting. 14. A demand quantity forecasting method for forecasting a demand quantity which fluctuates depending on weather conditions by using a neural network, wherein the hourly demand actual result data of each past time is preliminarily calculated before the learning of the neural network. Classify based on the maximum temperature, obtain the average of the hourly demand amount and the maximum temperature for each class, and as the input data of the input and output data to the neural network, the weather of the class, the average The maximum temperature is used as the output data for each hourly average demand of the class, and the actual values of the input / output data of all classes are used as the learning data, and the output of the neural network and the hourly average demand actual value match. As described above, the neural network is made to perform learning, and after the learning is completed, the forecasted weather and forecasted maximum temperature at the forecasted day when the demand forecast is performed are input to the neural network, and each time of the forecasted day is entered. A demand forecasting method characterized by forecasting hourly demand in a belt. 15. A demand forecasting method for forecasting a demand varying according to weather conditions by using a neural network, wherein the difference in weather between the current day and the previous day is used as input data of input / output data to / from the neural network. , Maximum temperature difference, sunshine time difference, and actual demand difference between the most recent weather day and the previous day's time zone, weather difference, maximum temperature difference, sunshine duration difference as output data Using the actual value of the input / output data for a certain period as learning data, the neural network is arranged so that the output of the neural network and the difference between the hourly demand actual values of the current day and the previous day match. After the learning, the neural network after learning finishes the demand forecast. The forecasted weather difference, forecasted maximum temperature difference, forecasted sunshine duration difference, and forecasted day, the same weather day and the previous day. Of each hour demand A demand amount forecasting method, characterized by inputting a record difference, a weather difference, a maximum temperature difference, and a sunshine duration difference, and predicting a time demand difference between the forecast day and the preceding day. 16. A demand forecasting method for forecasting a demand varying with weather conditions using a neural network, wherein the neural network is used to generate past hourly demand data based on the weather and maximum temperature of the day. Prepared for each class that has been divided into classes, and as the input data of the input / output data that is input / output to / from the neural network, the maximum temperature difference, the sunshine duration difference, and the recent same class date and the recent Using the actual difference in hourly demand amount, maximum temperature difference, and sunshine duration difference for each time zone of the same class day before the same class day as the output data, the difference in hourly demand amount for the current day and the previous same class day Using the actual value of the input / output data for a predetermined number of days as the learning data, the neural network is trained so that the output of the neural network and the difference in the actual demand value of the same day of the current day and the previous same class match. After the learning is completed, the neural network after the completion of learning predicts the demand amount, the predicted maximum temperature difference, the predicted sunshine duration difference, and the most recent same-class day and the most recent same-class day before the predicted day and the predicted day In front of the same class, enter the hourly demand amount difference, maximum temperature difference, sunshine duration and difference,
A demand forecasting method characterized by forecasting a difference in hourly demand in each time zone of the same day of the forecast day and the day before the forecast day. 17. A demand quantity forecasting method for forecasting a demand quantity fluctuating depending on weather conditions by using a neural network, wherein the weather, the maximum temperature, and the
On the day of the week, the actual demand data for each hour is classified, and the average of the hourly demands for the past specified number of days belonging to the same class is calculated as the forecasted value of the hourly demand. Direct fitting method The hourly demand in the band is predicted, and as the input data of the input / output data input / output to / from the neural network, the weather on the day, the maximum temperature, the sunshine time, and the hourly demand according to the prediction on the most recent weather day Quantity prediction error, maximum temperature, sunshine hours, using each hour demand forecast error by the prediction of the day as output data, the actual value of the input and output data of a certain period as learning data,
The neural network is made to perform learning so that the output of the neural network and the actual value of the prediction error of the time demand amount by the prediction match, and the demand prediction is made to the neural network after the learning is completed. Enter the forecast weather, forecast maximum temperature, forecast sunshine duration, and forecast error for each hour demand, forecast maximum temperature, and sunshine duration for the forecast date, and enter the forecast time on the forecast date. A demand forecasting method, comprising forecasting a forecast error of the demand forecast. 18. A demand quantity forecasting method for forecasting a demand quantity which fluctuates depending on a weather condition by using a neural network, wherein the actual demand value of each time zone in a predetermined period is previously calculated before learning of the neural network. The average is obtained in advance, and as the input data of the input / output data input / output to / from the neural network, the weather on the day, the maximum temperature, the sunshine time, and the actual time demand amount at each time of the latest weather day are calculated. The difference between the hourly average demand, the maximum temperature, and the sunshine duration are used as output data, and the difference between the hourly demand and the hourly average demand at each time of the day is used, and the actual value of the input / output data for a certain period of time. As learning data, the neural network is made to perform learning so that the output of the neural network and the difference between the actual demand value on the day and the average demand value on the day match, and the neural network after the learning is completed. In the circuit network, Enter the forecast weather, forecast maximum temperature, forecast sunshine duration, and the difference between the actual hourly demand and the hourly average demand in each time zone of the latest weather day, maximum temperature, and sunshine duration. Then, the demand forecasting method is characterized by forecasting the difference between the hourly demand and the hourly average demand in each time zone of the forecast date. 19. A demand forecasting method for forecasting a demand varying with weather conditions using a neural network, wherein the input data of input / output data to / from the neural network is the weather and maximum temperature of the day. , Sunshine hours, and the ratio of the hourly demand actual value to the daily demand for each time zone, the maximum temperature, and the sunshine duration of the most recent weather day as output data, the daily demand for the actual hourly demand value for each time zone of the day Using the actual value of the input / output data for a certain period of time as learning data by using the ratio to the amount, the output of the neural network and the ratio of the actual hourly demand value of the day to the daily demand amount are matched. , Forecasting weather, forecasted maximum temperature, forecasted sunshine duration, and hourly demand of the most recent weather day for which the neural network is made to perform learning and the neural network after the learning is performed is subjected to demand forecasting. Proven Input the ratio to the daily demand, the maximum temperature, and the sunshine hours, and output the ratio of the hourly demand to the daily demand in each time zone of the relevant forecast day, and predict the hourly demand based on the forecasted daily demand. A demand forecasting method characterized by: 20. In a demand forecasting method for forecasting a demand varying with weather conditions using a neural network, the weather and maximum temperature of the day are used as input data of input / output data to / from the neural network. , Sunshine hours, and the ratio of the actual demand value for each hour over the past specified number of days to the daily demand quantity, the weather, the maximum temperature, and the sunshine time are output data, and the daily demand quantity for the actual demand value for each time zone of the day is output. Using the ratio to the actual value of the input / output data for a certain period of time as learning data, the output of the neural network matches the ratio of the hourly demand actual value of the day to the daily demand, so that The neural network is made to perform learning, and the neural network after the learning is completed, the demand forecast is made. Forecast weather, forecast maximum temperature, forecast sunshine duration, and the day of demand for each hour between the predetermined number of days. demand Input the actual value of the ratio to the amount, the maximum temperature, and the sunshine hours, and output the ratio of the hourly demand amount to the daily demand amount in each time zone of the relevant forecast day, and calculate the hourly demand amount based on the forecast value of the daily demand amount. A demand forecasting method characterized by causing forecasting. 21. A demand forecasting method for forecasting a demand varying according to weather conditions using a neural network, wherein the neural network is used to generate past hourly demand data based on the weather and maximum temperature of the day. Prepared for each class that was divided into classes, and as the input data of the input and output data that is input to and output from the neural network, the maximum temperature of the day, the sunshine hours, and the actual time demand of each time zone of the same class day The ratio of the value to the daily demand, the maximum temperature, and the sunshine hours are used as output data, and the ratio of the actual value of the hourly demand amount for each time zone of the day to the daily demand is used, and the input for the specified number of days in the same class day Using the actual value of the output data as learning data, the neural network is learned so that the output of the neural network and the ratio of the actual hourly demand value of the day to the daily demand amount match. In the neural network after learning, the predicted maximum temperature, predicted sunshine duration, and the ratio of hourly demand to the daily demand of the latest similar class day are recorded in the neural network. A demand amount characterized by inputting a time and outputting the ratio of the time demand amount to the daily demand amount in each time zone of the relevant forecast day, and performing the forecast of the time demand amount based on the forecast value of the daily demand amount. Prediction method. 22. A demand quantity forecasting method for forecasting a demand quantity which fluctuates depending on weather conditions using a neural network, in advance of learning of the neural network, in advance of a time demand actual value of each time zone in a predetermined period. Classify ratio data to daily demand based on weather and maximum temperature,
The average of the maximum temperature and the ratio of the daily demand amount to the daily demand amount is calculated for each class, and the weather of the class and the average maximum temperature are used as the input data of the input and output data to and from the neural network. , Using the average of the ratio of the hourly demand actual value of the class to the daily demand as output data, the actual value of the input / output data of all classes as learning data, and the output of the neural network, and the class So that the average of the ratio of the hourly demand actual value to the daily demand matches, the neural network is made to perform learning, and the neural network after the learning is completed The feature is that the weather and predicted maximum temperature are input, the ratio of the hourly demand to the daily demand in each time zone of the forecast day is output, and the hourly demand is predicted based on the forecasted daily demand. Tosu Demand prediction method. 23. In a demand forecasting method for forecasting a demand varying with weather conditions using a neural network, the difference in weather between the current day and the day before is used as input data of input / output data to / from the neural network. , The maximum temperature difference, the sunshine duration difference, and the difference in the ratio of the hourly demand actual value to the daily demand amount for each of the most recent weather days and the day before the same day, the weather difference, the maximum temperature difference, and the sunshine duration difference as output data. Using the difference between the ratio of the hourly demand actual value to the daily demand on the current day and the previous day, the actual value of the input / output data for a certain period is used as learning data, and the output of the neural network and the current day and the previous day The neural network is made to perform learning so that the difference in the ratio of the hourly demand actual value to the daily demand is the same, and the neural network after the learning is completed has a forecast day and its forecast day Predicted weather the day before Input the difference, the predicted maximum temperature difference, the predicted sunshine duration difference, and the difference in the ratio of the hourly actual demand value to the daily demand amount on the same weather day and the previous day of the forecast day, the weather difference, the maximum temperature difference, and the sunshine duration difference. Then, the difference in the ratio of the hourly demand amount to the daily demand amount in each time zone of the current day and the previous day is output, and the hourly demand amount is predicted based on the predicted value of the daily demand amount. Demand forecast method. 24. A demand forecasting method for forecasting a demand varying according to weather conditions by using a neural network, wherein the neural network is used for past hourly demand data based on the weather and maximum temperature of the day. Prepared for each class that has been divided into classes, and as the input data of the input / output data that is input / output to / from the neural network, the maximum temperature difference, the sunshine duration difference, and the recent same class date and the recent Before the same class day, the difference in the ratio of the actual demand value to the daily demand amount for each time zone on the same class day, the maximum temperature difference, and the sunshine duration difference are used as output data for the hourly demand amount on the current day and the previous same class day. Using the difference of the ratio to the daily demand, the output of the neural network and the time demand of the same class day before the same class day as the learning data using the actual value of the input / output data for a predetermined number of days in the same class day in the past. The neural network is made to perform learning so that the difference in the ratio of the actual quantity value to the daily demand is the same, and the neural network after the learning is completed has a forecast day and forecast date for forecasting demand. Predicted maximum temperature difference on the same class day before,
Enter the forecast sunshine duration difference, the difference between the ratio of the hourly actual demand value to the daily demand amount for each of the most recent same-class day and the previous previous same-class day, the maximum temperature difference, and the sunshine duration difference, and make the prediction. A demand characterized by outputting the ratio of the hourly demand amount to the daily demand amount in each time zone of the same day before the day of the day and the forecast date and forecasting the hourly demand amount based on the forecast value of the daily demand amount. Quantity prediction method. 25. A demand quantity forecasting method for forecasting a demand quantity which fluctuates depending on weather conditions by using a neural network, wherein the weather, the maximum temperature, and the
The actual data of the ratio of the hourly demand amount to the daily demand amount is classified by day of the week, and the average of the ratio of the hourly demand amount to the daily demand amount during the past predetermined number of days belonging to the same class is averaged to obtain the daily demand amount day. By a prediction method similar to the direct fitting method with the predicted value of the ratio to the demand amount, the ratio of the time demand amount to the daily demand amount in each time zone is predicted, and the input / output data input / output to / from the neural network is predicted. As input data, the weather of the day, the maximum temperature, the sunshine time, and the prediction error of the ratio of each hourly demand amount to the daily demand amount by the forecast of the most recent same day, the maximum temperature, the sunshine time, the output date of the day Using the prediction error of the ratio of each hourly demand amount to the daily demand amount by prediction, the actual value of the input / output data in a certain period is used as learning data, and the output of the neural network and The neural network is made to perform learning so that the actual value of the prediction error of the ratio of the time demand to the daily demand due to the above prediction is matched, and the demand prediction is made to the neural network after the learning is completed. Enter the forecast weather, forecast maximum temperature, forecast sunshine time, and forecast error of the ratio of each hourly demand to the daily demand by the forecast on the latest weather day, maximum temperature, and sunshine time A demand forecasting method, characterized in that a forecast error of the forecast of the ratio of the hourly demand to the daily demand in each time zone of the day is output, and the hourly demand is forecast based on the forecasted value of the daily demand. . 26. A demand forecasting method for forecasting a demand varying according to a weather condition by using a neural network, wherein the hourly demand actual value of each time zone in a predetermined period is preliminarily set before learning of the neural network. The average of the ratio to the daily demand is obtained, and as the input data of the input / output data input / output to / from the neural network, the weather of the day, the maximum temperature,
Daylight hours, and the difference between the ratio of the hourly demand actual value to the daily demand amount and the average ratio of the hourly demand actual value to the daily demand amount, the maximum temperature, and the sunshine time as the output data Using the difference between the ratio of each hourly demand amount to the daily demand amount and the average ratio of each hourly demand amount actual value to the daily demand amount, the actual value of the input / output data for a certain period is used as learning data, and The output of the network,
The neural network is made to perform learning so that the difference between the ratio of the actual time demand value to the daily demand amount and the average ratio of the time demand amount to the daily demand amount is matched, and the neural network after the learning is completed. The forecasted weather, forecasted maximum temperature, forecasted sunshine duration, and the ratio of the actual demand value for each hour to the daily demand for the forecasted day and the actual demanded day for each hour on the same weather day Enter the difference of the average ratio to the demand amount, the maximum temperature, and the sunshine time, and enter the ratio of the hourly demand amount to the daily demand amount and the average ratio to the daily demand amount of each hourly demand amount in each time zone of the forecast day. A demand quantity forecasting method, which outputs a difference to forecast a time demand quantity based on a forecast value of a daily demand quantity. 27. Learning and prediction by the neural network are executed while sequentially changing learning parameters to find a value of the learning parameter that minimizes the average prediction error,
The learning and prediction by the neural network are performed using the value of the learning parameter.
6. The demand forecast method according to any one of 6. 28. The neural network performs learning by using, as learning data, a predetermined number of days of the input / output data excluding the days when the actual demand value is significantly different from the actual demand value before and after, and after the learning is completed. The demand forecasting method according to any one of claims 1 to 26, characterized in that the neural network of (1) is used to forecast a demand on a forecast day.