JPH09215192A - Method for predicting daily load curve - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the workload and improve the predicting accuracy in preparing a predictive model. SOLUTION: In a method for predicting the daily load curve, a computer predicts the net system energy demand on a specific day by using a neural network. The method is divided into a first step (SA1) in which the neural network is made to learn by using learning data composed of weather data at each prescribed time on each day during a period before the specific day and discrimination data on weekdays, Saturday, and holidays, and actual net system energy demand values at each prescribed time on each day and a second step (SA2) in which the neural network is made to successively predict the net system energy demand at each prescribed time on the specific day by inputting the weather data at each prescribed time on the specific day and the discrimination data on weekdays, Saturday, and holidays to the neural network in addition to the above-mentioned learning data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention automatically predicts a daily load curve on a system control computer or a general-purpose computer at a central power supply command station, a local power supply command station, a system control station, etc. in various power systems. Regarding the method.

[0002]

2. Description of the Related Art Forecasting of a daily load curve in a power system is often performed by the experience and intuitive knowledge of a skilled operator, and almost all of the tasks rely on manual work. As a normal prediction method, a skilled operator intuitively predicts by referring to a daily load curve and a daily maximum power prediction value on a day similar to the weather of the prediction target day obtained by the weather forecast.

As an example of automating such a forecasting work, a forecasting model is created for each time as shown in FIG. 8 by a statistical method typified by multiple regression analysis or a neural network, and the power demand for each time is calculated. A method has been proposed in which the daily load curve is predicted by individually obtaining

In the above-mentioned conventional prediction method, since the prediction model is individually prepared for each time, the amount of work and the required time for that are enormous. Further, the daily load curve is a curve showing a time series change of the power demand amount for one day, but in the method of individually predicting the power demand amount for each time, the prediction considering time-series elements is performed. I can't do it.

[0005]

While a large amount of specialized knowledge is required to operate a power system, the number of skilled operators who have this knowledge is decreasing in recent years. On the other hand, the daily load curve is the basis of power generation planning, which should be called the basis of grid operation, and there is a strong demand for improvement of its prediction accuracy and automation. Further, as described above, in the method of creating and predicting a prediction model every hour of the day, a huge amount of work and time are required to create the prediction model, and it is impossible to make a prediction in consideration of time series.

The present invention has been made in order to solve the above-mentioned problems, and does not depend on a skilled operator, and predicts a daily load curve in consideration of time series characteristics with a small number of prediction models, and when a prediction model is created. The present invention aims to provide a daily load curve prediction method that reduces the workload of the above and improves the prediction accuracy.

[0007]

In order to solve the above-mentioned problems, the invention according to claim 1 is a method for predicting a daily load curve of the amount of power demand on a prediction target day by a computer using a neural network. Let the neural network learn using the learning data consisting of meteorological data for each predetermined time on each day before the day, distinction data for weekdays, Saturdays, and holidays, and actual power demand for each predetermined time on each day. In addition to the first step and the learning data, the weather data for each predetermined time of the prediction target day, the data for distinguishing weekdays, Saturdays, and holidays are input to the neural network, and the prediction target is repeatedly recollected at each predetermined time. The second step is to sequentially predict the power demand amount for each predetermined time in the day. Here, as the neural network, it is desirable to use a recurrent network capable of prediction in consideration of time series.

According to a second aspect of the present invention, meteorological data for each predetermined time on each day in the period prior to the forecast target day, weekday / Saturday / holiday distinction data, lunch break distinction data, for each predetermined time on each day The first step of learning the neural network using the learning data consisting of the actual power demand values of the above, and in addition to the learning data, meteorological data for each predetermined time of the prediction target day, weekday / Saturday / holiday distinction data Second, by inputting distinction data for lunch break into the neural network and repeatedly recalling the data at predetermined time intervals to sequentially predict the power demand at predetermined time intervals on the prediction target day.
And steps.

According to the third aspect of the present invention, meteorological data for each predetermined time of each day in the period prior to the prediction target day, weather difference from the same time on the day before each day or the day one week before, A first step in which a neural network learns using learning data composed of demand power differences from the same time;
In addition to the learning data, weather data for each predetermined time of the prediction target day, weather difference data from the same day on the day before the prediction target day or one week before is input to the neural network, and the day before the prediction target day is input. Alternatively, a second step of repeatedly recalling the power demand difference from the same time on the day one week before every predetermined time to sequentially predict the power demand difference at each predetermined time on the prediction target day, and the day before the prediction target day. Or, it can be calculated by adding the second
The third step of adding the power demand difference predicted by the step to sequentially predict the power demand for each predetermined time on the prediction target day.

According to the invention of claim 4, the first step of creating standard data based on standard weather and power demand in the period to which the prediction target day belongs, and each day of the period prior to the prediction target day A second step of making the neural network learn using the meteorological data for each predetermined time, the meteorological difference data that is the difference from the meteorological standard data, and the learning data that is the demand power difference that is the difference from the electric power standard data; In addition to the learning data, the weather data for each predetermined time of the prediction target day is input to the neural network, and the demand power difference, which is the difference between the standard data of the power for each predetermined time on the prediction target day and (standard power), is calculated. The third step of sequentially predicting, and the demand power difference predicted by the third step is added to the standard power data, and the demand is calculated every predetermined time on the prediction target day. Sequentially competence, is made of a fourth step of predicting.

According to a fifth aspect of the present invention, the average weather data for each predetermined time in the weekly period before the forecast target date, the lunch break distinction data, and the average power demand record for each predetermined time in the weekly period A first step of learning the neural network using learning data consisting of values; and, in addition to the learning data, weather data and lunch break discrimination data for each predetermined time of the forecast target day are input to the neural network, A second step in which the average power demand for each predetermined time on the prediction target day is sequentially predicted by repeatedly recalling every hour, the average power demand predicted by the second step, and one day from the weekly period. The third step of sequentially predicting the demanded electric energy for each predetermined time on the prediction target day based on the actual demanded electric energy value of the period excluding It become one.

[0012] The invention according to claim 6 is learning comprising weather data of each day in a period prior to the target day of prediction, discrimination data of weekdays, Saturdays, and holidays, and an actual value of power demand for each predetermined time of each day. The first step of learning a multi-input multi-output neural network using data, and in addition to the learning data, weather data for one day of the prediction target day and weekday / Saturday / holiday discrimination data The second step is to collectively predict the demanded electric energy for each predetermined time on the prediction target day.

The invention according to claim 7 is the weather data of each day in the period before the forecast target day, the data of weekday / Saturday / holiday distinction, the actual value of the amount of power demand at each predetermined time of each day, and the maximum day. The first step of learning into a multi-input / multi-output neural network using learning data consisting of one or both of power demand and minimum daily demand, and maximum daily demand and minimum demand In addition to the second step of predicting one or both of the quantities and the learning data, weather data for one day of the target day of prediction, weekdays, Saturdays,
Data for holidays and one or both of the predicted maximum daily power demand and minimum predicted power demand are input to the neural network to remind them, and the predicted power demand for each predetermined time on the forecast target date is collectively predicted. The third step is to

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. First, an embodiment of the invention described in claim 1 will be described. As shown in FIG. 1, this embodiment includes a learning step (SA1) of a neural network and a recall step (SA2) of the neural network. Here, as the neural network, it is preferable to use a recurrent network in order to perform prediction in consideration of time series.

The details of each step will be described below. (1) Neural network learning step (SA
1) The data used for learning is, as input data, hourly meteorological data (temperature, humidity, etc.) for each day in the period before the prediction target date (for example, several days to several years before the prediction target date). , Weekday / Saturday / holiday distinction data (1,0, etc.)
Therefore, as the output data, the actual value of the demanded electric energy for each hour of each day is used. For example, assuming that the prediction target date is February 1, and when using the data for one month before that, the data from January 1 to January 31 is used.

A recurrent network is trained using these data to construct a prediction model capable of predicting the hourly demanded electric energy on the prediction target day.

As is well known, the recurrent network is a neural network which internally includes a feedback connection, and it is useful for grasping time series data by having the information of one step before in the network. The recurrent network model includes a Jordan model including an input layer, a hidden layer, an output layer, and a state layer, an Elman model including an input layer, an intermediate layer, an output layer, and a context layer, and a Willi including an input layer and an output layer.
An ams, Zipser model, etc. can be used.

An example of the learning algorithm will be briefly described. For example, in the Williams and Zipser model, the sensitivity of the output value at a certain time to the connection weight is sequentially obtained along the time axis, and the output error with respect to the connection weight is differentiated. And the learning is performed while correcting the connection weight in real time so as to minimize the output error (RTRL).

(2) Recall Step of Neural Network (SA2) For the learned recurrent network, from the day when learning is started (January 1 in the above example) to the prediction target day (also 2)
Meteorological data until the end of the month (on the first day of the month) (forecast target days are based on weather forecasts), weekday / Saturday / holiday distinction data, actual power consumption until the day before the forecast target day (also January 31) A value is input, and recollection of the recurrent network is repeated 24 times every hour to sequentially predict the hourly power demand on the prediction target day. The data used in this recall step is the same as the data used in the learning step, except for the weather data for the target day of the prediction and the weekday / Saturday / holiday distinction data.

The hourly power demand on the forecast target day is
It becomes the recollection value (RNN) of the recurrent network itself as in the following Expression 1. Note that t is the time of prediction on the day of prediction.

[0021]

[Equation 1] Demand power (t) = RNN (weather data (t), weekday / Saturday / holiday distinction data)

In the recurrent network, since the previous input data and state can be reflected in the output,
By sequentially predicting every hour as shown in
It is possible to make a prediction that reflects the state before time, that is, considering the time series. As a result, it is possible to obtain a daily load curve that takes into account the tendency of changes in power demand. Further, since it is possible to make a prediction by constructing one prediction model and repeatedly using it, it is possible to save the labor and time for constructing the model. For example, when it is desired to obtain predicted values every 15 minutes or every 30 minutes, data every 15 minutes or every 30 minutes may be used during learning.

Next, an embodiment of the invention described in claim 2 will be described. This embodiment is almost the same as the embodiment of the invention described in claim 1, except that the learning step (S
As data used for A1) and the recall step (SA2), distinction data (1, 0, etc.) for lunch break (12:00 to 13:00, especially 13:00) is added.

In general, the power consumption decreases during the lunch break. Therefore, by adding the distinction data of the lunch break for learning and recall, in addition to the correlation between the weather and the power demand, the power consumption during the day is reduced. It is possible to construct a prediction model that can be expressed well and to create a daily load curve that considers time series characteristics. In this case, the hourly power demand on the prediction target day is expressed by the following mathematical formula 2. According to this embodiment, there is an advantage that even power demand during lunch break, which is relatively difficult to predict, can be predicted with high accuracy.

[0025]

[Equation 2] Demand power (t) = RNN (weather data (t), weekday / Saturday / holiday distinction data, lunch break distinction data)

Next, an embodiment of the invention described in claim 3 will be described. As shown in FIG. 3, this embodiment includes a learning step (SB1) of the neural network, a recall step (SB2) of the neural network, and a prediction value calculating step (SB3). Here, as the neural network, it is preferable to use a recurrent network as in the previous case.

The details of each step will be described below. (1) Neural network learning step (SB
1) The data used for learning are, as input data, hourly meteorological data for each day in a period before the prediction target date (for example, several days to several years before the prediction target date), the previous day of each day. Or it is the weather difference from the same time one day ago,
As the output data, the demand power difference from the same reference time is used. In the above example, 1 for each day from January 1st to January 31st
Hourly weather data and the day before or 1 for each of these
Meteorological difference data and demand difference will be used based on the same time on the previous day of the week.

A recurrent network is trained using these data to construct a prediction model capable of predicting the demand power difference from the same time on the day before the prediction target day or the day one week before. In addition, when the weather difference data and the demand power difference from the previous day are used as the learning data, a model for predicting the demand power difference from the day before the prediction target day is constructed, and the learning data from the day one week ago is constructed. When the weather difference data and the demand power difference are used, a model for predicting the demand power difference from the day one week before the prediction target day is constructed. Here, the weather difference data may be digitized as a difference in temperature and humidity and used as input data.

(2) Recall step of the neural network (SB2) All the data used in the learning step in the learned recurrent network, that is, hourly meteorological data on each day from January 1st to January 31st. , The weather difference data and power demand difference from the same time on the day before each day or one week ago, and the weather data until the last time of February 1, which is the forecast target day (according to the weather forecast), Enter the weather difference data from the same day on the previous day (January 31) or one week ago (January 25) of the forecast target day, and enter the same day on the previous day of the forecast target day or one week ago. Recall the power demand difference from. This recall is repeated 24 times every 1 hour to predict the demand power difference for each hour.

(3) Prediction value calculation step (SB3) The recall value by the above-mentioned recall step is not the power demand at each time of the prediction target day, but the day before (January 31) or the day before one week ( Since it is the power demand difference from the same time (January 25), the power demand at each time of the prediction target day is predicted using the following Expression 3 or Expression 4.

[0031]

[Equation 3] Demand power amount (d, t) = RNN (weather data (t), weather difference data (t)) + demand power amount (d-1,
t)

[0032]

[Equation 4] Demand power amount (d, t) = RNN (weather data (t), weather difference data (t)) + demand power amount (d-7,
t)

In the equations 3 and 4, d indicates the forecast target date. In addition, when the formula 3 uses the weather difference data from the previous day (January 31st) and the power demand difference at the time of learning and recall, the formula 4 uses the weather difference data from the day one week ago (January 25th). , When using the demand power difference.

Here, when the criterion for determining the weather difference data and the demand power difference is set to the previous day, it reflects the change in the demand power amount for one continuous day and one day ago. When it is used as a reference, it reflects the amount of power demand on the most recent same day.

Also in this embodiment, as shown in FIG. 2, it is possible to carry out the prediction in consideration of the time series property, and it is possible to obtain the daily load curve in which the changing tendency of the power demand is taken into consideration.
In addition, since it is possible to make predictions using only a single prediction model, it is possible to save work effort and time for model building. The power demand on the prediction target day does not change drastically as compared with the nearest day, and the change amount is small when viewed from the whole. In this embodiment, the demand power amount itself is not recalled in the recall step of the neural network, but the demand power difference as a slight change is recalled, so that highly accurate prediction is possible.

Next, an embodiment of the invention described in claim 4 will be described. In this embodiment, as shown in FIG. 4, a standard data creation step (SC1) in which standard weather data of the month or season to which the prediction target day belongs and standard power demand are used.
And learning step of neural network (SC2)
And the step of recalling the neural network (SC
3) and a predicted value calculation step (SC4). Here, as the neural network, it is preferable to use a recurrent network as in the previous case.

The details of each step will be described below. (1) Standard data creation step (SC1) Hourly average weather data (average temperature, average) for a certain period prior to the forecast target day according to conditions such as weekdays / holidays or days of the week in which the forecast target day belongs Humidity) and average power demand are calculated and these data are used as standard data. By this processing, it is possible to obtain the hourly meteorological data and the amount of power demand as standard data on a standard day (standard day) from which a sudden abnormality factor is eliminated.

(2) Neural network learning step (SC2) The data used for learning is, as input data, each day of a period before the prediction target date (for example, several days to several years before the prediction target date). It is meteorological data for each hour and meteorological difference data (temperature difference, humidity difference) that is a difference from standard meteorological data. As output data, an hourly demand power difference for each day is used. Here, the demand power difference is the difference between the actual value of the demand power amount for each hour of each day and the standard data of power (standard power amount). In the above example, January 1 to January 31
The hourly meteorological data on each day of the day, the meteorological difference data and the demand power difference in these periods are used.

A recurrent network is trained using these input / output data to construct a prediction model capable of predicting the hourly demand power difference on the prediction target day.

(3) Neural Network Recall Step (SC3) All the data used in the learning step, that is, the hourly meteorological data on each day from January 1st to January 31st, in the learned recurrent network. Input the weather difference data and the demand power difference and the weather data until the end of February 1st, which is the forecast target date, and recall the demand power difference from the standard electricity amount every hour on the forecast target date. . This recall is repeated 24 times every 1 hour to predict the demand power difference for each hour.

(4) Prediction value calculation step (SC4) Since the recall value by the above-mentioned recall step is not the demand power amount at each time of the prediction target day but the demand power difference from the standard power amount, the following formula is used. 5 is used to predict the power demand at each time on the prediction target day.

[0042]

[Equation 5] Demand power amount (t) = RNN (weather data (t), weather difference data (t)) + standard power amount (t)

Also in this embodiment, as shown in FIG. 2, it is possible to make a prediction in consideration of the time series property, and it is possible to save the labor and time for constructing the model. Furthermore, by introducing the concept of standard data, it is possible to make predictions that exclude peculiarities such as sudden weather fluctuations and electric power demand fluctuations, and it is possible to make predictions that are highly universal.

Next, an embodiment of the invention described in claim 5 will be described. In this embodiment, as shown in FIG. 3, the learning step (SB1) of the neural network.
And the step of recalling the neural network (SB
2) and a predicted value calculation step (SB3). The main difference from the embodiment of the invention of claim 3 is that the weekly average weather data and average power demand before the target day of prediction and the lunch break distinction data are used in the learning step and the recall step. As the neural network, it is preferable to use a recurrent network as in the previous case.

The details of each step will be described below. (1) Neural network learning step (SB
1) For the data used for learning, as input data, hourly average weather data (average temperature, average humidity, etc.) for one week to multiple weeks (weekly) prior to the prediction target day, and lunch break distinction data are used. Also, as output data, the average power demand per hour on a weekly basis (for example, 1 at 10 am
Average of the amount of power demand for a week) is used. In the above example, for example, when data for one week before the prediction target date (February 1) is used, 1 of January 25 to January 31 is used.
The hourly average weather data and the lunch break distinction data, and the hourly average electricity demand during this period will be used.

By using the average data for one week or a plurality of weeks in this way, it is possible to eliminate fluctuations in the amount of power demand due to the day of the week, and the time-series data becomes simple and the data of the recurrent network can be obtained. Learning accuracy can be improved. By training these data in the recurrent network, a prediction model for predicting the average power demand for one week or a plurality of weeks on the prediction target day is constructed from the weather data and the lunch break discrimination data. The average meteorological data can be obtained from the following Equation 6.

[0047]

(Equation 6)

Here, the average weather data (d, t) is the average weather data for one week at d day t, and i is the number of days. The average power demand can also be calculated by the same idea.

(2) Neural Network Recall Step (SB2) All the data used in the learning step, that is, January 25 to January 31, are added to the learned recurrent network.
Enter the average weather data for each hour of the day and the data for distinguishing lunch breaks, and the average power demand for each hour of this period. Furthermore, in addition to these learning data, the weather data (by weather forecast) up to the final time of the prediction target day and the data for distinguishing lunch breaks are input, and the hourly average power demand is recalled. This recall 2 every 1 hour
It is repeated four times to predict the average power demand per hour. This average power demand can be obtained from the following Equation 7.

[0050]

[Equation 7] Average power demand (d, t) = RNN (average weather data (d, t), distinction data for lunch break)

Here, the average power demand (d, t) is the average power demand for one week on the d day (prediction target day) t, and i is the number of days.

(3) Predicted value calculation step (SB3) Since the recalled value in the recalled step described above is not the power demand at each time of the prediction target day, it is, for example, the average power demand at that time over one week. The following formula 8 is used to predict the power demand at each time on the prediction target day.

[0053]

(Equation 8)

As a result, it is possible to obtain the amount of power demand at each time on the prediction target day. In this embodiment,
Since the input and output data at the time of learning and recall are given as average values, the predicted amount of power demand can be suppressed by the effects of sudden weather fluctuations. When using a prediction model that reflects time-series characteristics, good prediction accuracy can be obtained due to time-series characteristics, but when peculiar data is included in the middle of the time series, it may adversely affect subsequent predictions. .

However, this embodiment is characterized by the time series data input to the neural network as in the embodiment of the invention described in claim 4. That is, by using the deviation from the average value so as to suppress the adverse effect of the peculiar data while reflecting the time-series property, the prediction accuracy is improved and stable prediction is possible.

Next, an embodiment of the invention described in claim 6 will be described. As shown in FIG. 5, this embodiment includes a learning step (SD1) of a neural network and a recall step (SD2) by the neural network. In the recall step, meteorological data for one day to be predicted ( By inputting weather forecast data) and weekday / Saturday / holiday distinction data (1, 0, etc.), the hourly demand power amount is collectively predicted. As the neural network, it is preferable to use a recurrent network as in the previous case.

The details of each step will be described below. (1) Learning step of neural network (SD
1) The data used for learning is, as input data, meteorological data (temperature, humidity, etc.) on each day in a period before the forecast target date (for example, several days to several years before the forecast target date), weekdays and Saturdays. -This is the data that distinguishes holidays, and the actual value of the amount of power demand is used as output data. Here, the meteorological data does not necessarily have to be data for every fixed time, but may be the highest and lowest temperatures and the same humidity for each day, or the meteorological data for a specific time. Note that the output data uses the amount of power demand for each hour.

In learning, one day's worth of data is collectively input to the network, and learning is performed so that the hourly demanded electric energy for one day of the prediction target day can be collectively predicted and output. The recurrent network used for this learning is multi-input multi-output (24 outputs) as shown in FIG. 7, and a specific output terminal specially outputs the power demand at a specific time. It can be performed.

(2) Neural Network Recall Step (SD2) For the learned recurrent network, the meteorological data from the day when learning is started to the end of the forecast target day (the forecast target day is based on the weather forecast), weekdays, Input the Saturday / holiday distinction data and the actual value of the amount of power demand up to the day before the forecast target date, and collectively predict the hourly power demand for one day of the forecast target day by recalling once . The data used here is the same as the data used in the learning step except for the data for the prediction target date.

In this embodiment, even when data cannot be obtained hourly such as a weather forecast, only the characteristic data of one day is used to collectively collect the hourly power demand for one day of the forecast target day. It is possible to predict efficiently. Also, in a multi-input multi-output recurrent network, one output terminal corresponds to each time and makes a dedicated prediction, so the work and time for building a prediction model can be reduced, and highly accurate prediction can be performed. .

Next, an embodiment of the invention described in claim 7 will be described. In this embodiment, as shown in FIG. 6, a learning step (SE1) of a neural network and a daily maximum / minimum power prediction step (SE2) by another method,
It consists of a recall step (SE3) of a neural network. Again, it is desirable to use a recurrent network.

The details of each step will be described below. (1) Neural network learning step (SE
1) The data used for learning is, as input data, meteorological data (temperature, humidity, etc.) on each day in a period before the prediction target date (for example, several days to several years before the prediction target date), weekdays and Saturdays. -Use the holiday distinction data, the maximum daily power demand, the minimum daily power demand, or both actual values. As the output data, the actual value of the power demand on each day is used.
The meteorological data does not necessarily have to be data for every fixed time, but may be the highest and lowest temperatures, the same humidity for each day, and meteorological data for a specific time. Note that the output data uses the amount of power demand for each hour.

In learning, as in the embodiment of the invention of claim 6, one day's worth of data is collectively input to the network, and the hourly power demand for one day of the forecast target day is collectively entered. Learn to be able to predict and output. The recurrent network used for this learning is also multi-input multi-output (24 outputs) as shown in FIG.

(2) Daily maximum / minimum power prediction step (S
E2) This step is a step of predicting the maximum daily power demand, the minimum daily power demand, or both by using another method. As a concrete prediction method, there are a method of making a prediction using an appropriate mathematical expression, a method of making a prediction based on the experience and intuition of a skilled operator, and a method of making a prediction using another neural network.

(3) Neural Network Recall Step (SE3) The learned recurrent network has the data used in the learning step, the meteorological data until the final time of the forecast target day (according to the weather forecast), and the maximum daily power demand or By inputting the minimum daily power demand or both, the power demand per hour for one day of the forecast target day is collectively predicted by one recall.

According to this embodiment, the same effect as that of the embodiment of the invention of claim 6 can be obtained. Furthermore, although it depends on the electric power company, in general, a skilled operator or the like generally predicts the maximum daily power demand and then predicts the daily load curve. That is, the daily load curve is predicted from the maximum daily power demand and the weather on that day, and the present embodiment also meets such a current situation, and effectively uses the prediction result of the maximum daily power demand. However, it can be said that this is a highly accurate prediction method that takes into consideration the time series characteristics and the correlation with the weather.

[0067]

As described above, according to the inventions of claims 1 to 7, the time series can be obtained by one prediction model by a neural network (in particular, a recurrent network) without depending on a skilled operator. It is possible to accurately predict the amount of power demand at each time on the prediction target day while considering the property. Particularly, in the invention according to claim 1,
Compared with the inventions of the other claims, the number of data used for learning and the number of processing steps for prediction can be relatively small.

According to the second aspect of the present invention, it is possible to satisfactorily predict power demand during lunch break, which is difficult to predict, and it is possible to contribute to improvement in prediction accuracy. Claim 3
According to the described invention, the demand power difference as a change from the reference date is predicted, so that highly accurate prediction is possible.

According to the invention described in claim 4 or the invention described in claim 5, by introducing the deviation from the standard data or the average value in learning and recall, the adverse effect of the peculiar data in the time-series data is exerted. It is possible to suppress the error and make highly accurate and stable prediction.

According to the invention described in claim 6, even when the weather data for each time cannot be prepared as in the weather forecast, 1
It is possible to predict all of the power demands for one day of the forecast target day only with the characteristic data of the day. In particular, if a multi-input multi-output type recurrent network is used, each output terminal will be in charge of prediction of each time, which enables highly accurate prediction.

According to the invention of claim 7, in addition to the effect of the invention of claim 6, it is possible to utilize the prediction result such as the maximum daily power demand in the same manner as the current prediction method, and it is possible for a skilled operator to It is possible to maintain consistency with the forecast.

[Brief description of drawings]

FIG. 1 is a flow chart showing an embodiment of the invention described in claims 1 and 2.

FIG. 2 is a diagram conceptually showing a prediction method according to an embodiment of the invention described in claims 1 to 5.

FIG. 3 is a flow chart showing an embodiment of the invention described in claims 3 and 5.

FIG. 4 is a flowchart showing an embodiment of the invention described in claim 4;

FIG. 5 is a flowchart showing an embodiment of the invention described in claim 6;

FIG. 6 is a flowchart showing an embodiment of the invention described in claim 7.

FIG. 7 is a diagram conceptually showing a prediction method according to an embodiment of the invention described in claims 6 and 7.

FIG. 8 is a diagram conceptually showing a conventional prediction method.

Claims (7)

[Claims] 1. A method of predicting a daily load curve of the amount of power demand on a forecast target day by using a neural network by a computer, comprising: meteorological data at a predetermined time on each day of the period prior to the forecast target day; A first step of learning to a neural network using learning data composed of Saturday / holiday distinction data and actual demand power amount for each predetermined time of each day; and, in addition to the learning data, a predetermined time of a prediction target day The second step of inputting the weather data for each time and the data for distinguishing weekdays, Saturdays, and holidays into the neural network, and repeatedly recalling them at predetermined time intervals to sequentially predict the power demand for each predetermined time on the prediction target day. A method for predicting a daily load curve, comprising: 2. A method of predicting a daily load curve of power demand on a forecast target day using a neural network by a computer, wherein meteorological data for each predetermined time of each day in a period before the forecast target day, weekdays, A first step of learning to a neural network using learning data composed of Saturday / holiday distinction data, lunch break distinction data, and actual power demand at each predetermined time of each day; and, in addition to the learning data, a prediction Input the meteorological data for each predetermined time of the target day, the weekday / Saturday / holiday distinction data, and the lunch break distinction data into the neural network, and remind them repeatedly at the predetermined time to predict the power demand for the predetermined time on the target day. And a second step of sequentially predicting, and a daily load curve predicting method comprising: 3. A method of predicting a daily load curve of power demand on a forecast target day using a neural network by a computer, comprising: meteorological data for each predetermined time on each day in a period prior to the forecast target day; In addition to the learning data, a first step in which a neural network is trained using learning data consisting of a weather difference from the same time on the day before the day or a week ago and a demand power difference from the same time; The weather data for each predetermined time of the forecast target day, the weather difference data from the same day on the previous day of the forecast target day or on the day of the week before the forecast target day is input to the neural network, and the day before the forecast target day or the day before the week A second step of repeatedly recalling the power demand difference from the same time every predetermined time to sequentially predict the power demand difference at each predetermined time on the prediction target day, and a prediction pair Predict the power demand for each predetermined time on the target day by adding the power demand difference predicted in the second step to the power demand actual value at the same time on the day before the elephant day or the week before A daily load curve prediction method comprising the following third step: 4. A method of predicting a daily load curve of the demand power amount of a forecast target day using a neural network by a computer, creating standard data based on standard weather and demand power amount of a period to which the forecast target date belongs. Power demand, which is the difference between the first step and the meteorological difference data that is the difference between the meteorological data and the standard data of the meteorological data for each predetermined time of the period before the target day for prediction The second step of learning the neural network by using the learning data consisting of the difference; and, in addition to the learning data, the meteorological data for each predetermined time on the prediction target day is input to the neural network, and for each predetermined time on the prediction target day. The third step of sequentially predicting the demand power difference, which is the difference from the standard power data, and the third step in the standard power data. Sequential demand power amount for each predetermined time in the prediction target day by adding more demands power difference expected, fourth step and, daily load curves prediction method characterized by comprising the to be predicted. 5. A method of predicting a daily load curve of the amount of power demand on a forecast target day using a neural network by a computer, wherein average weather data for each predetermined time in a weekly period prior to the forecast target day, lunch break Distinction data, a first step of learning by a neural network using learning data composed of average power demand actual values for each predetermined time of a weekly period, and in addition to the learning data, a predetermined time of a prediction target day A second step of inputting weather data for each time and data for distinguishing lunch breaks into the neural network and causing it to be repeatedly recalled at predetermined time intervals to sequentially predict the average power demand for each predetermined time on the prediction target day; Based on the average power demand predicted by the step and the actual power demand for the period excluding one day from the weekly period And a third step of sequentially predicting the amount of power demand for each predetermined time on the prediction target day, and a daily load curve prediction method comprising: 6. A method of predicting a daily load curve of the amount of power demand on a forecast target day by using a neural network by a computer, the meteorological data of each day in a period prior to the forecast target day, weekdays,
A first step of learning a multi-input / multi-output neural network using learning data consisting of Saturday / holiday distinction data and actual demand power amount for each predetermined time of each day; and, in addition to the learning data, a prediction The second step of inputting weather data for one day of the target day and data for distinguishing weekdays, Saturdays, and holidays into the neural network and recalling them to collectively predict the amount of power demand for each predetermined time on the prediction target day And a method for predicting a daily load curve, which comprises: 7. A method for predicting a daily load curve of the amount of power demand on a forecast target day by using a neural network by a computer, the meteorological data of each day in a period before the forecast target day, weekdays,
A multi-input / multi-output neural network using learning data consisting of Saturday / holiday distinction data, actual power demand at a predetermined time on each day, and / or maximum daily demand / minimum demand power A first step of learning, a second step of predicting one or both of the maximum daily power demand and the minimum daily power demand of the forecast target day, and one day of the forecast target day in addition to the learning data Of the weather data, weekday / Saturday / holiday distinction data, and predicted one or both of the maximum daily power demand and minimum daily power demand are recalled by inputting them to the neural network, A daily load curve prediction method comprising: a third step of collectively predicting the amount of power demand;