JP2795997B2 - Water distribution prediction device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an improvement in a water distribution predicting apparatus for predicting daily water distribution from the viewpoint of ensuring planned operation of water supply facilities.

[0002]

2. Description of the Related Art Generally, a water supply facility takes a route of taking raw water from a dam-type reservoir or a river in a mountainous area, and transporting the raw water taken at a water intake to a water purification plant through a headrace. Therefore, in such a water supply facility, there is a large delay in introducing raw water from the intake plant to the water treatment plant, and the water treatment plant uses a water purification process of coagulation and sedimentation, so there is a large time delay until distribution. For this reason, the time from water intake to water distribution changes greatly sequentially. For this reason, it is difficult to feedback-control the water intake at the water intake and the water supply at the water purification plant based on the fluctuation of the water distribution at the final end. Therefore, the amount of water distribution on the day is predicted in advance,
It is important to make daily operation plans for water intakes, water purification plants and water distribution plants. By the way, the water distribution varies depending on social living conditions such as 1) weather conditions, temperature, snowfall, sunshine hours, etc., 2) unusual days such as days of the week, holidays, May holidays, summer holidays, new year holidays, etc. I do.

Therefore, conventionally, a method of predicting daily water distribution from past actual data for each living condition such as a day of the week while considering such demands and influences, and a statistical model of the daily water distribution with consideration of weather conditions. (For example, when the temperature is high, the daily water distribution is large) to predict the daily water distribution. Further, the daily distribution is performed by repeating learning while changing the weighting coefficient using one neural network. A method for estimating the amount of water has been considered.

[0004]

However, the former method of predicting living conditions such as the day of the week or the prediction method taking into account weather conditions does not have a learning ability to use actual data obtained daily as a teacher signal. There are problems such as the inability to follow long-term structural changes due to construction and the like, and low prediction accuracy because the prediction method is based on one of the conditions.

[0005] On the other hand, in the latter prediction method using a neural network, learning results in consideration of fluctuations of various conditions are recorded in one neural network throughout the year, so that, for example, the temperature, which is one of the weather conditions, is determined. Very wide range of maximum and minimum temperature of the year,
As a result, the prediction accuracy of the daily water distribution becomes low, and the daily water distribution cannot be accurately predicted.

[0006] The present invention has been made in view of the above circumstances, and it is possible to accurately predict the amount of daily water distribution regardless of the annual fluctuation of various conditions, thereby ensuring economical and stable water supply in water supply facilities. It is an object to provide a water distribution amount prediction device.

[0007]

In order to solve the above-mentioned problems, the present invention performs a required classification process based on weather information and water distribution amount information, and includes a unique day for each month. Day of the week information that does not include the day, weather information (for example, weather,
A deviation data from the average maximum temperature, deviation from the average minimum temperature) and actual data classification means to classify the water distribution amount for each month are provided, and a neural network for each month is provided in advance. By inputting weekday information / weather information and water distribution amount for each month including the anomalous day of each month classified as described above, and excluding the anomalous day, and learning by changing the weighting coefficient, Providing a monthly forecast model creating means for creating a daily water distribution forecast model for each month, further selecting a neural network of the corresponding month from among the forecasted models according to the corresponding month, The water distribution amount provided with a daily water distribution amount prediction means for predicting the daily water distribution amount on the day by inputting the day of the week information, weather information, and the daily water distribution amount of the previous day including no specific day. It is a measuring apparatus.

[0008]

Therefore, according to the present invention, by taking the above-described means, the actual data classifying means takes in weather information and day information including a special day as required, and, in addition to day information for each month, For example, weather, deviation from average maximum temperature,
The information is classified into meteorological information such as deviation from the average minimum temperature and stored in a file. Similarly, the water distribution amount from the water distribution flow rate detector is taken in, classified into monthly water distribution information for each month, and stored in a file.

If the day of the week information, weather information, and water distribution amount for each month are classified into files as described above, the monthly water distribution amount is predicted by each month prediction model creating means using the classified information. Here, a plurality of neural networks are prepared in advance, a neural network is selected for each month, information classified for the neural network for the month is input, and the daily water distribution amount of the day is given as a teacher signal, and Then, a prediction model is created for each month while repeating learning with variable weighting factors.

After the prediction model is created in this way,
The daily water distribution forecasting means selects the neural network of the corresponding month, inputs the day of the week including the special day, weather information and the daily water distribution of the previous day to the neural network, and predicts the daily water distribution of the day. By doing so, the daily water distribution is predicted with high accuracy.

[0011]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a configuration diagram showing one embodiment of the device of the present invention. In this device, a water distribution flow rate detector 4 is installed in the middle of distribution of clean water from a reservoir 1 through a water distribution pipe 2 to a water distribution network 3 by natural flow, that is, in the water distribution pipe 2. The distribution flow rate detected by 4 is taken in by an input device (process interface) 5 and
The day of the week, weather information, and the amount of daily water distribution of the previous day are taken in by the input device 6 such as a keyboard and a mouse to predict the amount of daily water distribution of the day. Means 10, monthly prediction model creation means 20, daily water distribution amount prediction means 30, and the like are provided.

The performance data classifying means 10 includes the day of the week and weather information (for example, temperature) taken from the input device 6.
Is stored on the basis of the day / weather information in the result file 11, for example, if the date is 30 days, a predetermined process is performed for each day from 1 to 30 days, and the process is performed on a monthly basis ( The day / weather classification calculating means 12 for classifying into the day / meteorological data of each month) and the day / weather classification file 13 for storing the day / meteorological data of each month classified by the classifying means 12 are included. Is provided.

Further, the performance data classifying means 10 stores a water distribution amount file 14 for storing the water distribution amount taken in by the input device 5 and, based on the actual water distribution amount of the water distribution amount file 14, for example, on the 30th of the month. If there is, the predetermined processing is performed every day for 1 to 30 days, and the distribution amount classifying means 15 for classifying into the distribution amount results on a monthly basis (every month). A distribution amount classification file 16 for storing the classified distribution amount data for each month is provided. The water distribution amount classification file 16 also stores the water distribution amount of the day, which becomes a teacher signal (output).

Next, each month prediction model creation means 20
There are provided 12 sets of neural networks corresponding to two months, and after selecting a neural network of a corresponding month on the basis of calendar (month) data, performance classification from the two classification files 13 and 16 is performed on the neural network. Learning function 2 for inputting data and teacher signal to obtain optimum output while changing weighting coefficient repeatedly
By learning at step 1, a prediction model 22 for each month is created.

Further, the daily water distribution amount predicting means 30 selects a neural network of a month corresponding to calendar (month) data from the prediction model 21, and takes in the selected neural network 31 from the input device 6. The day of the week, weather information, and the amount of daily water distribution of the previous day are input, and the normalized daily water distribution of the current day is predicted and output. Numeral 32 denotes a signal conversion means for converting the normalized daily water distribution on the day into an actual daily water distribution. The daily water distribution converted by the signal converter 32 is displayed on the screen of the display device 7 such as a CRT. Is done. Next, the operation of the above-described apparatus will be described separately for the actual data classifying unit 10, the monthly forecast model creating unit 20, and the daily water distribution amount predicting unit 30. A. About the operation of the performance data classification means 10

The performance data classifying means 10 stores the day of the week and the weather information sequentially taken by the input device 6 in the day of the week and the weather performance file 11, and the distribution amount detected by the distribution flow rate detector 4 to the input device 5. And stored in the actual water distribution amount file 14.

Here, the day / weather classification calculating means 12 performs a predetermined classification process described later based on the day / weather information data stored in the file 11 and classifies the data into day / weather data for each month. The distribution amount calculation means 15 performs the same predetermined classification (normalization) processing based on the distribution amount actual data stored in the file 14 to categorize the distribution amount actual data for each month. Then, the water distribution amount of the day, which is to be a teacher signal (output), is stored in the classification file 16 by performing a predetermined classification (normalization) process. Therefore, first, the classification process of the day of the week, the weather, and the amount of water distribution on the previous day for input to the neural network of each month prediction model creating means 20 will be described. A-1 Information on Day of the Week The information on the day of the week is converted into the following 7-bit 0, 1 pattern and stored in the classification file 13. Sunday → 1,000,000 Monday → 0100000 Tuesday → 0010000 Wednesday → 0001000 Thursday → 0000100 Friday → 0000010 Saturday → 0000001 A-2 Meteorological Information A-2a. For weather information in the weather forecast,

The expression is divided into the morning and the afternoon of the previous day and the day. The expression method is divided into, for example, three types of fine, cloudy, and rainy, and converted into the following 3-bit 0 and 1 patterns, respectively, and stored in the classification file 13. Fine → 100 Cloudy → 010 Rain → 001 A-2b. The information about the temperature is divided into the maximum temperature and the minimum temperature, and the deviation from the average maximum temperature and the average minimum temperature is obtained and classified.

(A). Information on maximum temperature: First, the average maximum temperature and the variance of the maximum temperature are calculated from the actual temperature data. Now, assuming that the maximum temperature result on the i day is θmax (i) [° C], the average maximum temperature is the average θmax [°
C]

[0020]

(Equation 1) Can be represented by Further, the average θmax [° C]
The variance σθmax [° C] of the maximum temperature using

[0021]

(Equation 2) It is calculated from the following arithmetic expression. Here, n is the number of data.

Next, the maximum temperature record θmax (i) of the i-th day
[° C], mean θmax [° C] and variance σ of maximum temperature
From θmax [° C], the average θma is calculated based on the following equation (3).
The deviation (tilde θmax) from x (i) is obtained.

[0023]

(Equation 3)

(B). Information about minimum temperature… Classifies information about minimum temperature in the same way as maximum temperature. i
Assuming that the daily minimum temperature result is θmin (i) [° C], the average minimum temperature is the average θmin [° C],

[0025]

(Equation 4) Can be represented by Furthermore, this average θmin [° C]
The variance σθmin [° C] of the minimum temperature using

[0026]

(Equation 5) It can be obtained from the following equation. Here, n is the number of data.

Further, the minimum temperature result θmin (i) of the i-th day
[° C], mean θmin [° C] and variance σ of minimum temperature
From θmin [° C], the average θ is calculated based on the following equation (6).
The deviation (tilde θmin) from min (i) is obtained.

[0028]

(Equation 6) A-3 Information on water distribution on the previous day

Information on the water distribution amount of the previous day is obtained by the following method.
Classify. Here, the actual daily water distribution amount on (i-1) day [m
^Three / Day] is Q (i-1),
The daily water distribution amount Qyester (i) can be expressed by Qyester (i) = {Q (i−1) −Qmin} / (Qmax−Qmin) (7) Where Qmax is the maximum daily water distribution
[M^Three / Day] and Qmin is the minimum daily water distribution [m^Three / Day], n
Is the number of data. A-4 Classification of prediction target information

Here, the information is classified by the following method from the information on the characteristic of the water distribution amount on the day (i), which is a teacher signal of the neural network constituting each month prediction model creating means 20. In this case as well as the water distribution amount of the previous day, i
Actual daily water distribution [m ³ / Day] is Q (i), and the normalized daily water supply amount Qlearn (i) on the i-th day is: Qlearn (i) = {Q (i) -Qmin} / (Qmax-Qmin) 8)

Therefore, the day of the week / weather classification calculating means 12 calculates, for example, the day of the week, the weather (fine,
Cloudy, rainy), deviation from average θmax (i) (tilde θmax)
And classification information such as deviation (tilde θmin) from the average θmin (i) is obtained and stored in the day / weather classification file 13,
On the other hand, the daily water distribution classification calculating means 15 obtains the daily water distribution Qyester as the teacher signal in addition to the previous day's water distribution Qyester for the number of days in the month, and stores it in the water distribution classification file 16. B. About operation of each month prediction model creation means 20

The month predictive model creating means 20 selects a neural network corresponding to the month of the calendar, and stores the classified days of the week and the weather-related information stored in the files 13 and 16 in the input layer neurons of the neural network. In addition to the information, the daily water distribution amount on the previous day is input, and the daily water distribution amount on the current day is classified as a teacher signal for the output of the output layer neurons. While the learning is repeated by the learning function 21 while varying, a prediction model 22 for each month having an optimal weighting factor is created.

In this learning, a back propagation method is used. The back propagation method is a learning method in which a network error propagates back from an output layer to an input layer in a neural network having a hierarchical structure. The learning procedure will be described below. The suffix for each month is omitted because the formula becomes complicated.

Step 1: On the day of the week, weather-related information is input to the input layer of the neural network, and the intermediate layer and the output layer calculate according to the following neuron model. The output Hj of the jth neuron in the hidden layer is

[0035]

(Equation 7)

It is obtained from the following equation. Here, I ₁ is the output of the i-th neuron in the input layer, W _1j is the weight coefficient of the i-th neuron in the input layer and the j-th neuron in the intermediate layer, 1 is the number of input layers, m
Is the number of hidden layers, and f () is the threshold function of the hidden layer. On the other hand, the output ok of the k-th neuron in the output layer is

[0037]

(Equation 8) Obtained from Here, W _jk is a weight coefficient between the j-th neuron in the _hidden layer and the k-th neuron in the output layer, m is the number of hidden layers,
n is the number of output layers.

Step 2: Learning is performed by modifying the weighting coefficient of the network so as to minimize the sum of the square error between the output ok of the kth neuron in the output layer and the teacher signal yk of the kth neuron in the output layer. The learning of the weight coefficient between the intermediate layer and the output layer is to calculate △ W _{jk in the} following equation and correct W _jk . W _jk (t-1) = W _jk (t) + △ W _jk (t) (11) ΔW _jk (t) = − ε · d _k (t) · H _j (t) (12) _{) d k (t) = ok} (t) -yk (t) ...... (13) where, t is the number of times of learning in the above equation, epsilon determines the size of one modified parameter, d _k is the output layer It is an error.
Next, the learning of the weight coefficient between the input layer and the output layer is performed by calculating △ W _ik based on the following equation and correcting W _ik .

[0039]

(Equation 9) In the above equation, _dj is the back propagation error of the intermediate layer, and f () is f
This is the derivative of (). Furthermore, the convergence of learning is expedited by reducing vibration using the following equation. ΔW _jk (t) = − ε · d _k (t) · H _j (t) + α · △ W _ik (t−1) (17) ΔW _ij (t) = − ε · dj (t) I _i (t) + α · △ W _ij (t-1) (18) In the above equation, the preceding term is the correction amount, and the following term excluding α is the previous weight. α is a parameter for stability.

Therefore, each month prediction model creation means 20 creates a prediction model 22 by memorizing the weighting factors obtained by the above learning in different neural networks for each month. Therefore, it is possible to obtain a prediction model 22 composed of 12 neural networks for each month. C. Operation of daily water distribution amount prediction means 30

The daily water distribution amount prediction means 30 selects a neural network 31 corresponding to the month of the calendar from the prediction model 22 for each month. When the day of the week, the weather information, and the daily water distribution amount of the previous day are input to the neural network 31, the neural network 31 has a weighting factor set by learning in advance. Daily water distribution can be predicted. Since the predicted daily water distribution amount on the day is normalized, it is converted into the actual daily water distribution amount by the signal conversion means 32 and displayed on the display device 7.

Therefore, according to the configuration of the embodiment described above, the day of the week, the actual weather, and the actual water distribution amount from the distribution water flow rate detector 4 are converted into classification data by performing required classification processing. A file can be created for each month so that it can be used in an optimal state for the neural network.
In addition, each month prediction model creating means 20 prepares an independent neural network for each month, selects a neural network of a corresponding month, inputs the filed classification data, and repeats learning to obtain an optimal weight. Since the prediction model 22 having the coefficient is created, the fluctuation range of the maximum and minimum temperatures in a monthly unit is small, and it is possible to predict the daily water distribution amount with high accuracy, and to learn by incorporating various conditions comprehensively. The prediction model 22 can be created without having such a complicated configuration.

Furthermore, a prediction model 22 is created using the daily water distribution obtained on a daily basis as a teacher signal, and the daily water distribution is predicted using the prediction model 22. Can sufficiently follow structural changes.

FIG. 2 is a diagram showing the results of identifying the daily water distribution. The illustrated solid line indicates the actual water distribution amount, and the illustrated dotted line indicates the predicted water distribution amount. The predicted water distribution amount is almost the same as the actual water distribution amount. It can be seen that the prediction accuracy is very high.

Although the day information is stored in the file 11 in the above embodiment, the day information may be directly converted into the classification file 13 and stored without being stored in the file 11. In addition, the apparatus for predicting the daily water distribution amount by learning the actual water distribution amount on the day has been described.For example, the number of neurons in the middle layer of the neural network for each month may be changed variously during model creation. Build the model,
What is necessary is just to select a model with a high validity rate. Further, the present apparatus is applied to the prediction of daily water distribution, but can be applied to other prediction targets such as power prediction as long as a neural network is constructed based on features such as a day of the week and weather. Further, by adding an input of a specific day (May consecutive holidays, Bon holidays, etc.) to the input layer of the neural network, more accurate prediction can be performed. Furthermore, although the daily water distribution is output on the day, this device has a causal relationship with the input and output, for example, when the temperature is high and clear, the water distribution is large.
When the temperature is low and rain is low, it can be easily applied to the case where the amount of water distribution is small. In addition, the present invention can be implemented with various modifications without departing from the scope of the invention.

[0046]

As described above, according to the present invention, a prediction model is created using the classification information for each month and the neural network for each month, and the neural network for the month is selected to distribute daily data. Since the amount of water is predicted, even if the annual fluctuation of various conditions is large, it is possible to accurately predict the amount of daily water distribution, and to provide a water distribution amount predicting apparatus capable of sufficiently securing economical and stable water supply in a water supply facility.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing one embodiment of a water distribution amount prediction device according to the present invention.

FIG. 2 is a diagram showing a relationship between an actual water distribution amount and a predicted water distribution amount of a daily water distribution amount.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Distribution reservoir, 2 ... Distribution pipe, 4 ... Distribution flow rate detector, 5,6
... input device, 7 ... display device, 10 ... actual data classification means, 11 ... day and weather actual calculation file, 12 ... day and weather classification arithmetic means, 13 ... day and weather classification file, 14 ...
Water distribution amount actual file, 15 ... water distribution amount classification calculating means, 16
... Water distribution amount classification file, 20... Monthly prediction model creating means, 21... Learning function, 22.
... means for predicting daily water distribution, 31 ... neural network for the corresponding month, 32 ... data conversion means.

Claims (1)

(57) [Claims] 1. A water distribution amount prediction device for predicting a water distribution amount to be distributed from a water supply facility, performs a required classification process based on weather actual information and water distribution amount information, and performs day information, weather information and monthly information. Actual data classifying means for classifying water distribution amounts, and a neural network for each month are provided in advance, and the day of the week information and weather information for each month and the water distribution amount for each month classified for the neural network of the corresponding month. And a monthly prediction model creating means for creating a daily water distribution amount prediction model for each month by learning by changing the weighting coefficient. Select the neural network of the day, and input the day of the week information, weather information of the day, and the amount of water distribution of the previous day to this neural network to predict the amount of daily water distribution of the day. A water distribution amount prediction device comprising: a water amount prediction unit.