Wind power generation prediction device, wind power generation prediction method, and wind power generation prediction program

Abstract

To provide a wind power generation prediction device, a wind power generation prediction method, and a wind power generation prediction program for efficiently and accurately predicting the amount of power generated by a wind power generator.SOLUTION: A machine learning model 17 has a configuration in which sequence translation models of pieces of time-series data are connected in series. An information acquisition unit acquires wind speed prediction data at a plurality of points in a predetermined area where a plurality of wind power generators are installed in various places, and actual result data of the total amount of power generated by the plurality of wind power generators at the time of prediction of the wind speed prediction data. A learning execution unit causes a machine learning model 17 to perform learning, with the wind speed prediction data and the actual result data of the total power generation amount as learning data. An execution unit inputs the wind speed prediction data to the learned machine learning model 17 to predict the total amount of power generated by the plurality of wind power generators.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a wind power generation prediction device, a wind power generation prediction method, and a wind power generation prediction program. [Background technology]

[0002] In the power supply business of the electric utility industry, the balance between electricity generation and consumption must be simultaneous and equal in quantity. Simultaneous and equal in quantity means that the amount of electricity generated, i.e., supply, and the amount of electricity consumed, i.e., demand, are equal at the same time. If the amount of electricity generated and consumed does not match, the frequency of the electricity will be disrupted, the quality of the electricity will deteriorate, and it will become difficult to supply electricity normally. Adjustments to simultaneous and equal amounts are generally made for each power balancing group and for a wide area that aggregates multiple balancing groups. For this reason, a large number of different types of generators will be deployed within a wide area.

[0003] Since electricity cannot be stored on a large scale, achieving a simultaneous balance requires controlling the amount of electricity produced by adjusting the output of power plants so that it matches the amount of electricity consumed. Generators have different characteristics depending on their type: thermal power, which is easy to adjust the amount of electricity produced; nuclear and hydroelectric power, which produce a constant amount of electricity; and solar and wind power, which are difficult to adjust and require prediction. Supply and demand adjustment personnel work tirelessly, predicting the amount of electricity generated by solar and wind power plants and the amount of electricity consumed, and adjusting the number of operating and stopped generators and the amount of electricity generated to keep the difference between the amount generated and the amount consumed within a certain range.

[0004] With the increase in renewable energy, accurate predictions of power generation from solar, wind, etc. are required to achieve simultaneous balancing. However, if these predictions are incorrect, unexpected costs such as startup costs, fuel costs, and electricity purchase costs will be incurred. Therefore, various technologies have been proposed to predict power generation amounts.

[0005] For example, a technology has been proposed that uses a recurrent neural network with a mechanism for feeding back a layer mapped to a first-order lag hidden layer as a context layer to predict wind speeds several hours in advance based on past and current weather data, thereby predicting the power generated by wind power generation facilities. Another technology has been proposed that uses a prediction model that learns the correlation between actual wind speeds in a specified area and actual power generated by wind power generators, thereby predicting the power generated by wind power generators from predicted wind speeds within a specified area. Another technology has been proposed that uses meteorological analysis data, which is an analysis of wind observation data, wind turbine measurement data, and weather forecast data, in a recurrent neural network to predict wind conditions for a first period and a shorter second period, and then predicts wind power generation based on the prediction results.

[0006] Other proposed technologies include using LSTM (Long Short Term Memory) to predict wind speeds and converting the predicted wind speed values using a power curve to calculate predicted power output. Another proposed technology uses past wind power output as input, adds the electricity price in the hour-ahead market as a weight for the loss function, and uses LSTM to calculate wind power output in the short term. Another proposed technology predicts wind speeds based on 10-minute average wind speed, 10-minute maximum wind speed, and temperature, and then calculates wind power output from the prediction results. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Japanese Patent Application Laid-Open No. 2007-56686 [Patent Document 2] Japanese Patent Publication No. 2021-182319 [Patent Document 3] Japanese Patent Application Publication No. 2023-107068 [Non-patent literature]

[0008] [Non-Patent Document 1] Maki Fujiwara, Bo-Tong Chen, and Shoji Kawasaki, "Study on Short-Term Wind Power Output Forecasting Using LSTM with Surrounding Area Weather Data," 2023 National Conference of the Institute of Electrical Engineers of Japan [Non-patent document 2] Lu Huimin, Yu Fujimoto, and Yasuhiro Hayashi, "A Study on Price-Aware Short-Term Forecasting of Wind Power Generation for Hour-Ahead Market Energy Trading," 2023 National Conference of the Institute of Electrical Engineers of Japan [Non-patent document 3] Yuancheng Li and Akira Sugawara, "Power Stabilization of Wind Power and Pumped Storage Power Generation Systems by Wind Speed Prediction Using LSTM," 2022 Electrical Engineering Conference of the Institute of Electrical Engineers of Japan, Power and Energy Division Summary of the Invention [Problem to be solved by the invention]

[0009] However, in conventional wind power generation forecasts, a machine learning model is created for each wind power generator, and the output of each wind power generator is predicted using the trained machine learning model. The results are then used to predict the output of each wind power generator on a regional basis. This method requires detailed information about each wind power generator, as well as various usage information about the wind power generator, such as its start-up, expansion, maintenance, and decommissioning. Obtaining all of this usage information is not easy in a wide area where numerous wind power generators are installed. Furthermore, machine learning models for each wind power generator require maintenance to adapt them to actual operating conditions, which creates a significant workload. Furthermore, if wind power generators without a prediction system exist within the wide area being forecast, it is difficult to accurately predict the power generation of the entire area. As such, conventional forecasting methods make it difficult to efficiently and accurately predict the power generation of wind power generators in a wide area.

[0010] Furthermore, the power output of a wind turbine is primarily determined by wind, which changes continuously over time. Therefore, treating wind information as time-series data can improve prediction accuracy. Furthermore, when making predictions using wind information, wind conditions prior to the time of prediction are also an important factor. In this regard, a technology for predicting power output using a recurrent neural network with a feedback mechanism treats wind as time-series data, but does not take into account conditions prior to the time of prediction, making it difficult to accurately predict the power output of a wind turbine. Furthermore, a technology for predicting power output from predicted wind speed using a prediction model that learns the correlation between actual wind speed and actual power output also does not take into account changes in wind conditions prior to the time of prediction, making it difficult to accurately predict the power output of a wind turbine. Furthermore, a technology for predicting power output from predicted wind conditions for a first period and a second period obtained by applying meteorological analysis data to a recurrent neural network also does not take into account changes in wind conditions prior to the time of prediction, making it difficult to accurately predict the power output of a wind turbine.

[0011] Furthermore, a technology that calculates power output predictions by converting wind speed predictions using LSTM with a power curve only predicts wind power generation from the same generator at a single site, making it difficult to accurately predict power generation in a wide area where many different types of wind turbines are installed. Furthermore, a technology that calculates wind power generation using LSTM by adding hourly market electricity prices as weights to past wind power generation is a prediction based solely on electricity prices, and is not based on wind speed, etc., making it difficult to accurately predict power generation that is appropriate for the environment. Furthermore, a technology that calculates wind power generation from wind speed predictions based on 10-minute average wind speed, etc., assumes the installation of the same wind turbine at a single site, making it difficult to accurately predict power generation in a wide area where many different types of wind turbines are installed.

[0012] The present invention has been made in consideration of the above, and aims to provide a wind power generation prediction device, a wind power generation prediction method, and a wind power generation prediction program that efficiently and accurately predict the amount of power generated by a wind power generator. [Means for solving the problem]

[0013] In the present invention, the machine learning model has a configuration in which series transformation models for time-series data are connected in series. The information acquisition unit acquires forecast data for wind speeds at multiple locations in a predetermined area where multiple wind power generators are installed in various locations, and actual data for the total power generation amount of the multiple wind power generators at the time of prediction of the wind speed forecast data. The learning execution unit trains the machine learning model using the forecast data for wind speeds and the actual data for the total power generation amount as learning data. The execution unit inputs the forecast data for wind speeds into the trained machine learning model, and predicts the total power generation amount of the multiple wind power generators. [Effects of the Invention]

[0014] According to the present invention, it is possible to efficiently and accurately predict the amount of power generated by a wind power generator. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is a block diagram of a wind power generation prediction device. [Figure 2] FIG. 2 is a diagram showing an example of a time schedule for processing to predict the amount of power generated by wind power generation in a wide area. [Figure 3] FIG. 3 is a diagram showing an overview of the sequence to sequence model. [Figure 4] FIG. 4 is a diagram illustrating an overview of LSTM. [Figure 5] FIG. 5 is a diagram illustrating a wind power generation prediction algorithm according to the first embodiment. [Figure 6] FIG. 6 is a diagram showing details of the machine learning model for wind power generation prediction. [Figure 7] FIG. 7 is a diagram illustrating the ReLU function. [Figure 8] FIG. 8 is a diagram showing an example of prediction results and performance data of the amount of power generated by wind power generation in a wide area. [Figure 9]FIG. 9 is a flowchart of the machine learning process performed by the wind power generation prediction device according to the first embodiment. [Figure 10] FIG. 10 is a flowchart of the power generation amount prediction process performed by the wind power generation prediction device according to the first embodiment. [Figure 11] FIG. 11 is a diagram showing a comparison of prediction accuracy. [Figure 12] FIG. 12 is a diagram showing the daily transition of the predicted results of the amount of power generation using a plurality of prediction methods and the actual value of the amount of power generation. [Figure 13] FIG. 13 is a diagram illustrating a wind power generation prediction algorithm according to the second embodiment. [Figure 14] FIG. 14 is a diagram showing an example of the configuration of the coupling layer. [Figure 15] FIG. 15 is a flowchart of the machine learning process performed by the wind power generation prediction device according to the third embodiment. [Figure 16] FIG. 16 is a hardware configuration diagram of the wind power generation prediction device. DETAILED DESCRIPTION OF THE INVENTION

[0016] The following describes in detail embodiments of a wind power generation prediction device, a wind power generation prediction method, and a wind power generation prediction program disclosed in the present invention with reference to the accompanying drawings. Note that the wind power generation prediction device, the wind power generation prediction method, and the wind power generation prediction program disclosed in the present invention are not limited to the following embodiments. [Example]

[0017] FIG. 1 is a block diagram of a wind power generation prediction device. The wind power generation prediction device 1 predicts the amount of wind power generation in a wide area using forecast data on wind speeds at various points in a predetermined wide area and actual data on the amount of power generation by wind turbines installed in the wide area. An example of a wide area is the Tohoku area (Aomori, Iwate, Akita, Miyagi, Yamagata, Fukushima, Niigata) in the electric power industry. This wide area is an example of a "predetermined area." The amount of power generation by wind turbines in a wide area is the total amount of power generation by multiple wind turbines installed in the wide area.

[0018] In the first embodiment, the wind power prediction device 1 is a device that predicts the amount of wind power generated in a wide area. Hereinafter, the day targeted for prediction will be referred to as the target prediction day. The wind power prediction device 1 predicts the amount of wind power generated in a wide area using time-series continuous wind speed data, for example, predicting the amount of wind power generated at each time every 30 minutes. In the first embodiment, the wind power prediction device 1 performs predictions every 30 minutes from 9:00 on the day before the target prediction day until just before midnight on the target prediction day, which is 39 hours later, to obtain prediction results for the amount of wind power generated in the wide area every 30 minutes for the entire target prediction day. In this way, the wind power prediction device 1 of the first embodiment actually performs predictions for the target prediction day and a predetermined time immediately before the target prediction day, and obtains prediction results for the target prediction day from among the prediction results. In the following description, the first time of the time slot that is actually the target of prediction will be referred to as the reference time. In the first embodiment, the reference time is 9:00.

[0019] The following describes in detail the wind power generation prediction device 1. As shown in Fig. 1, the wind power generation prediction device 1 has an information collection unit 11, a data storage unit 12, an aggregation and organization unit 13, a prediction execution unit 14, a performance evaluation unit 15, a learning execution unit 16, a machine learning model 17, and a result output unit 18. The wind power generation prediction device 1 has two operation phases: a learning phase in which the machine learning model 17 is made to learn, and a prediction phase in which the trained machine learning model 17 is used to predict unknown future power generation amounts of wind power in a wide area.

[0020] The information collection unit 11 scrapes various data used for learning or prediction in each of the learning phase and the prediction phase. For example, the information collection unit 11 collects wind speed forecast data from the weather information system 2 at each location in a wide area from 24 hours before the reference time until midnight on the prediction target date. In Example 1, the information collection unit 11 collects wind speeds at 33 locations in the wide area. As for the acquisition timing, for example, the information collection unit 11 collects hourly forecast data up to midnight on the prediction target date every hour until two days before the prediction target date, and then collects data every 30 minutes from midnight on the previous day. The information collection unit 11 then stores the acquired wind speed forecast data for the prediction target date at each location in the data storage unit 12.

[0021] If there is missing information in the collected wind speed prediction data, the information collecting unit 11 supplements it using wind speed prediction data acquired by the information collecting unit 11 in subsequent rounds. The information collecting unit 11 leaves the missing data in the newly collected wind speed prediction data among the collected wind speed prediction data, and overwrites the collected wind speed prediction data with the newly collected prediction data. By repeatedly acquiring and overwriting wind speed prediction data in this way, the information collecting unit 11 can collect wind speed prediction data for every hour up to midnight on the prediction target day without omission.

[0022] Here, the reason for repeatedly collecting wind speed prediction data is to avoid missing any data, and the information collection unit 11 does not need to repeat collection once all of the hourly wind speed prediction data up to midnight on the prediction target day is collected. If there is no chance of missing any prediction data, the information collection unit 11 may be configured to collect wind speed prediction data once, just before making the prediction.

[0023] Fig. 2 is a diagram showing an example of a time schedule for the process of predicting the amount of power generated by wind power generation in a wide area. Fig. 2 shows an example in which the prediction target date is January 2nd, and the amount of power generated by wind power generation in a wide area is predicted every 30 minutes from 9:00 on January 1st, the previous day, as the reference time, until just before midnight on January 2nd.

[0024] The information collecting unit 11 collects hourly wind speed forecast data at least up to midnight on January 2nd, and every hour on December 31st. Furthermore, from midnight to 8:30 on January 1st, the information collecting unit 11 collects wind speed forecast data every 30 minutes. This allows the information collecting unit 11 to collect wind speed forecast data for each hourly time exemplified at timings 101 and 102.

[0025] Furthermore, in the learning phase, the information collection unit 11 collects performance data on the amount of power generated by wind power generation in the wide area from the wind power generator power generation performance database 3. The wind power generator power generation performance database 3 is, for example, a database provided by each electric power company. The information collection unit 11 then stores the acquired performance data on the amount of power generated by wind power generation in the wide area in the data storage unit 12.

[0026] Returning to FIG. 1, the explanation will be continued. The aggregation and organization unit 13 performs preprocessing, including aggregation and organization, on the data stored in the data storage unit 12. For example, the aggregation and organization unit 13 rearranges the forecast data of wind speed and the actual data of power generation amount in a wide area, and complements missing data. If there is missing data, the aggregation and organization unit 13 may complement the missing data using data similar to the missing data, or may exclude data for the day when there is missing data from the data used for learning and prediction. The aggregation and organization unit 13 also assigns a predetermined file name to each piece of data, organizes the storage location, and generates a list indicating the storage location.

[0027] For example, in the example shown in FIG. 2, at time 105, the aggregation and sorting unit 13 rearranges and interpolates the forecasted wind speed data and the actual data of the power generation amount of wind power generation in the wide area collected up to time 102.

[0028] Returning to FIG. 1, the explanation will continue. Thereafter, in the learning phase, the aggregation and organization unit 13 outputs to the learning execution unit 16 the forecast data of wind speed for each point for each hour from 24 hours before the reference time used for learning to midnight on the target prediction day. The aggregation and organization unit 13 also outputs to the learning execution unit 16 the actual data of the amount of power generated by wind power generation in the wide area for each hour from the reference time to midnight on the target prediction day. Here, the combination of the forecast data of wind speed and the actual data of the amount of power generated by wind power generation in the wide area becomes the learning data, and among these, the actual data of the amount of power generated by wind power generation in the wide area becomes the correct label. A plurality of data from different dates is used for the learning data.

[0029] Furthermore, in the prediction phase, the aggregation and organization unit 13 outputs hourly wind speed prediction data for each location from 24 hours before the reference time to midnight on the prediction target date to the learning execution unit 16. This wind speed prediction data becomes prediction data to be used for prediction.

[0030] The machine learning model 17 is a prediction model that predicts the amount of power generated by wind power generation in a wide area using predicted wind speed data as an input. The machine learning model 17 according to the first embodiment uses a sequence-to-sequence model that receives time-series data as an input and obtains time-series data as an output. The sequence-to-sequence model has a configuration in which two recurrent neural networks (RNNs) that perform sequence conversion between sequence data are connected in series.

[0031] Fig. 3 is a diagram showing an overview of a sequence to sequence model. As shown in Fig. 3, the sequence to sequence model 300 has an encoder 301, which is an RNN on the input side, and a decoder 302, which is an RNN on the output side. The encoder 301 has an RNN 311, which is a hidden layer in which multiple RNNs are connected in series. The decoder 302 has an RNN 312, which is a hidden layer in which multiple RNNs are connected in series.

[0032] For example, wind speed, which is time-series data, is input over time to each of the serially arranged RNNs 311 in the encoder 301 in accordance with the order of the arrangement. As a result, the output of each RNN 311 is propagated, and an eigenlength vector h representing the characteristics of the wind speed, which is the input time-series data, is output from the last RNN 311 in the series. Next, the eigenlength vector h is input to the first RNN 312 in the series, and the wind speed, which is time-series data, is input over time in accordance with the order of the serially arranged RNNs 312 in the decoder 302. As a result, the output of each RNN 312 is propagated, and a predicted value of the power generation amount is output from the last RNN 312 in the series.

[0033] In particular, here, the machine learning model 17 uses a model equipped with a Long Short Term Memory (LSTM) network, which is a type of RNN. LSTM networks are explicitly designed to avoid problems of exploding gradients and vanishing gradients, and have the ability to efficiently learn long-term dependencies in sequence data. Hereinafter, the LSTM network will be simply referred to as "LSTM."

[0034] As described above, the machine learning model 17 has a main algorithm that handles time-series data using a sequence-to-sequence model. Here, RNNs that handle time-series data are deep neural networks that handle the time axis, making them prone to the vanishing gradient problem. Furthermore, when given input that is "low relevance at the present time but will become more relevant in the future," there is an input weight conflict problem—whether to increase or decrease the weight. Therefore, the LSTM was devised to solve these problems. LSTM introduces a mechanism called the LSTM block that appropriately retains time-series information within the network. The LSTM block primarily includes cells for retaining errors and gates for retaining or erasing necessary information at the required timing. The main algorithm of the machine learning model 17 uses LSTM in the RNN portion of the sequence-to-sequence model.

[0035] Figure 4 is a diagram showing an overview of an LSTM block. The LSTM block 320 in Figure 4 corresponds to the RNNs 311 and 312 in Figure 3. The LSTM block 320 has memory cells that hold information to be stored for long-term storage. σ321 is a forget gate. σ322 is an input gate. σ323 is an output gate.

[0036] Here, t is the step of the LSTM block 320, that is, the number of stages from the top of the serially connected LSTM blocks 320. t is the short-term memory of the RNN. t is a state variable for long-term storage of short-term information. The memory cell has a cell state C t is stored as internal memory. The storage cell is in cell state C t Then, the previous short-term memory h that you want to retain in the next step t Weight and store x t is an observation vector, and in this embodiment, is information obtained from predicted data of wind speed.

[0037] The forget gate σ321 is f t Outputs f t is the cell state C t The input gate σ322 is a value that determines how much information to discard from i t Output i t is the new information to be stored in the memory cell. ~ t (C t The value of the input gate σ321 is the new information stored in the memory cell. The output gate σ323 is t Outputs o t is the weighting factor of new information output from the memory cell. t ,f t ,o t ,C ~ t ,C t and h t is expressed by the following formula (1).

[0038]

number

[0039] where W is the weight matrix connecting the previous LSTM block 320 in series with the current LSTM 320, and U is the weight matrix added to the input to the LSTM 320. In this way, LSTMs are explicitly designed to avoid the problem of long-range dependencies.

[0040] Learning conditions are set in advance for the machine learning model 17. However, after the machine learning model 17 is given appropriate learning conditions and trained, the learning conditions may be set again.

[0041] 5 is a diagram illustrating a wind power generation prediction algorithm according to Example 1. The machine learning model 17 according to Example 1 includes an encoder 171 that is an input-side LSTM and a decoder 172 that is an output-side LSTM.

[0042] Wind speed #1 in FIG. 5 represents the predicted wind speed data at 9:00 a.m. on the day before the prediction target date for 33 locations in a wide area. Wind speeds #2 to #79 represent the predicted wind speed data for every 30 minutes up to 39 hours after 9:00 a.m. on the day before the prediction target date. Wind speeds #0 to #-47 represent the predicted wind speed data for every 30 minutes within the 24 hours before 9:00 a.m. Here, standardized values, as described below, are used as the predicted wind speed data input to machine learning model 17.

[0043] The encoder 171 has 48 combinations of fully connected layers and LSTM encoders 710 connected thereto, and the LSTM encoders 710 are connected in series. For each combination, wind speeds #-47 to #0 over the past 24 hours, obtained as performance data, are input to the fully connected layer in the order they are arranged in series. For example, wind speed #-47 is input to the first LSTM encoder 710 connected in series via the fully connected layer. Values output from each LSTM encoder 710 other than the last one are input to the next LSTM encoder 710 connected in series. Each LSTM encoder 710 other than the first one connected in series receives predicted wind speed data via the fully connected layer, and also receives the value output from the previous LSTM encoder 710 connected in series. The last LSTM encoder 710 connected in series outputs a latent feature 173.

[0044] The LSTM encoder 710 can obtain latent features 173 that incorporate all of the time-series data, wind speeds #-47 to #0, by sequentially propagating the output obtained from the wind speed at each time. Conceptually, these latent features 173 can be said to be information summarizing the state of wind speed over the most recent 24 hours.

[0045] The decoder 172 has 79 combinations of a front-stage fully connected layer, an LSTM decoder 720, and a rear-stage fully connected layer, and each LSTM decoder 720 is connected in series. For each combination, wind speeds #1 to #79 obtained as prediction data for 39 hours from 9:00 a.m. onward are input to the front-stage fully connected layer in the order they are arranged in series.

[0046] Wind speed #1 is input to the fully connected layer preceding the first combination connected in series, and its output is input to the LSTM decoder 720 of that combination. In addition, the latent feature 173 output from the encoder 171 is input to the LSTM decoder 720 of the first combination connected in series. The value output from the LSTM decoder 720 of the first combination connected in series using these as input is input to the fully connected layer following the first LSTM decoder 720 connected in series. The fully connected layer following the first combination connected in series uses the value output from the LSTM decoder 720 as input, and outputs power generation E1, which is the predicted value of wind power generation in the wide area at 9:00 a.m. on the day before the prediction target date in the wide area.

[0047] The fully connected layer at the previous stage of each combination other than the first in the series receives the corresponding wind speed prediction data, and the output is input to the LSTM decoder 720 for that combination. Each LSTM decoder 720 in each combination other than the first receives the value output from the previous LSTM decoder 720. The value output from each LSTM decoder 720 in each combination other than the first is then input to the fully connected layer at the subsequent stage. The value output from each LSTM decoder 720 other than the last in the series is then input to the next LSTM decoder 720. The fully connected layer at the subsequent stage of each combination other than the first also receives the value output from the LSTM decoder 720 and outputs a prediction value for the amount of wind power generated in the wide area at the corresponding time. For example, the fully connected layer at the subsequent stage of the second combination outputs power generation E2, which is the prediction value for the amount of wind power generated in the wide area at 9:30 a.m. on the day before the target prediction date. The fully connected layer after the final combination connected in series outputs power generation E79, which is the predicted value of wind power generation in the wide area at 23:30 on the target forecast day.

[0048] In this way, each LSTM decoder 720 propagates the wind speed at each time along with information on the wind speed state for the previous 24 hours in sequence, thereby making it possible to make a prediction that takes into account the time-series changes in wind speed from the previous 24 hours to midnight on the prediction target day. Therefore, the decoder 172 can take into account the most recent wind speed state in chronological order and the wind speed on the prediction target day in chronological order when predicting the amount of power generation, improving the accuracy of the time-series prediction of the amount of power generation from wind power generation over a wide area.

[0049] Fig. 6 is a diagram showing details of a machine learning model for wind power generation prediction. The machine learning model 17 has, for example, the structure shown in Fig. 6. Fig. 6 shows an example in which the Tohoku area is targeted as a wide area.

[0050] The conditions used in the machine learning model 17 in Figure 6 are as follows: The explanatory variables are wind speed forecast data for 33 locations in the Tohoku area. The objective variable is the actual wind power generation output. The algorithm is LSTM. There are 256 hidden layers. The activation function is the ReIU function. The ReIU function is expressed by the following formula (2) and Figure 7. Figure 7 is a diagram showing the ReLU function. In Figure 7, the horizontal axis represents the variable x in the ReLU function, and the vertical axis represents f(x).

[0051]

number

[0052] In this machine learning model 17, time-series data on predicted wind speeds at each location is input to an input layer corresponding to each of 33 predetermined locations in the Tohoku area. Then, via 256 hidden layers, predicted values for wind power generation in the Tohoku area are output from the output layer of machine learning model 17.

[0053] In the first embodiment, a case has been described in which one prediction model is used as the machine learning model 17. However, for example, a plurality of prediction models may be used in which the parameters of the prediction models of the same algorithm are finely adjusted.

[0054] Continuing the explanation by returning to Fig. 1, in the learning phase, the learning execution unit 16 receives input of learning data from the aggregation and organization unit 13, including wind speed prediction data and actual data on the amount of power generated by wind power generation in a wide area.

[0055] Next, the learning execution unit 16 standardizes the wind speed prediction data. For example, the learning execution unit 16 standardizes the obtained wind speed prediction data. In this standardization, the learning execution unit 16 performs an averaging process on the wind speed prediction data, which is data for each hour, to generate wind speed prediction data for each 30 minutes.

[0056] The learning execution unit 16 then inputs the standardized wind speed prediction data to the machine learning model 17. Specifically, the learning execution unit 16 inputs the wind speed prediction data for every 30 minutes within the 24 hours before 9:00 on the day before the prediction target day to the encoder 171 of the machine learning model 17. The learning execution unit 16 also inputs the wind speed prediction data for every 30 minutes for the 39 hours from 9:00 on the day before the prediction target day until midnight on the prediction target day to the decoder 172 of the machine learning model 17.

[0057] The learning execution unit 16 then acquires the forecast results for wind power generation in the wide area every 30 minutes for 39 hours from 9:00 a.m. on the day before the target prediction date to midnight on the target prediction date, which are output from the decoder 172. The learning execution unit 16 then compares the acquired forecast results for wind power generation in the wide area with actual data on wind power generation in the wide area, and adjusts parameters to reduce the loss function. When the learning termination conditions, such as the number of repeated epochs and the upper limit time, are met, the learning execution unit 16 terminates the training of the machine learning model 17. As a result, the learning execution unit 16 generates a trained machine learning model 17.

[0058] In the first embodiment, the learning execution unit 16 divides the learning data into a plurality of groups and uses mini-batch learning to train the above-described machine learning model 17. By using mini-batch learning, the learning execution unit 16 can cause the machine learning model 17 to perform stable learning at a high calculation speed.

[0059] In the prediction phase, the prediction execution unit 14 receives input of prediction data including wind speed prediction data from the aggregation and organization unit 13. Similar to the learning execution unit 16, the prediction execution unit 14 standardizes the wind speed prediction data.

[0060] Next, the prediction execution unit 14 inputs the standardized wind speed prediction data to the trained machine learning model 17. Specifically, the prediction execution unit 14 inputs the wind speed prediction data for every 30 minutes within the 24 hours before 9:00 on the day before the prediction target day to the encoder 171 of the trained machine learning model 17. Furthermore, the learning execution unit 16 inputs the wind speed prediction data for every 30 minutes for the 39 hours from 9:00 on the day before the prediction target day to midnight on the prediction target day to the decoder 172 of the trained machine learning model 17.

[0061] Thereafter, the prediction execution unit 14 acquires the predicted values of the amount of wind power generated in the wide area for every 30 minutes for 39 hours from 9:00 on the day before the target prediction day to midnight on the target prediction day, which are output from the decoder 172. The prediction execution unit 14 then extracts the predicted values for every 30 minutes for 24 hours on the target prediction day from the acquired predicted values, and sets these as the predicted results of the amount of wind power generated in the wide area on the target prediction day.

[0062] Thereafter, the prediction execution unit 14 stores the prediction result of the wind power generation amount in the wide area on the obtained prediction target day in a predetermined location in the data storage unit 12. The prediction execution unit 14 also outputs the prediction result of the wind power generation amount in the wide area on the obtained prediction target day to the result output unit 18.

[0063] For example, in the case of FIG. 2, the prediction execution unit 14 standardizes the wind speed prediction data acquired before 9:00 and inputs it to the trained machine learning model 17. Then, the prediction execution unit 14 acquires the prediction results of the amount of wind power generation in the wide area every 30 minutes from 9:00 on January 1st to midnight on the prediction target day, which is 39 hours later, as illustrated by times 111 to 117, output from the machine learning model 17. The prediction execution unit 14 outputs the prediction results from time 116, which is midnight on the day after the prediction execution timing 120 in FIG. 2, to the result output unit 18 as the prediction result of the amount of wind power generation in the wide area on January 2nd, which is the prediction target day.

[0064] Returning to Figure 1, the explanation will continue. The performance evaluation unit 15 acquires the prediction result of the amount of power generated by wind power generation in the wide area on the day before the prediction target date and the performance data of the amount of power generated by wind power generation in the wide area on the day before the prediction target date from the data storage unit 12. Next, the performance evaluation unit 15 compares the prediction result of the amount of power generated by wind power generation in the wide area on the day before the prediction target date with the performance data of the amount of power generated by wind power generation in the wide area on the day before the prediction target date, to evaluate the performance. Thereafter, the performance evaluation unit 15 outputs the result of the performance evaluation to the result output unit 18.

[0065] For example, the performance evaluation unit 15 generates a graph showing the predicted and actual values of wind power generation in the wide area on the day before the target prediction date, and a graph showing the predicted and actual values of wind power generation in the wide area on the day before the target prediction date up to the reference time. The performance evaluation unit 15 also generates a graph showing the predicted values of wind power generation in the wide area on the target prediction date. Furthermore, the performance evaluation unit 15 calculates the accuracy of the prediction for the day before the target prediction date as the performance evaluation. The performance evaluation unit 15 then creates a table showing the results of the performance evaluation using the calculated accuracy of the prediction. In this way, the performance evaluation unit 15 creates graphs and tables as information showing the results of the performance evaluation.

[0066] 2, when prediction is performed at timing 120, the data storage unit 12 has already stored performance data for the period before timing 120. Therefore, the performance evaluation unit 15 evaluates the performance for December 31, which is the day before the target prediction date. The performance evaluation unit 15 then outputs the result of the performance evaluation for December 31 to the result output unit 18.

[0067] FIG. 8 is a diagram showing an example of a graph including forecast results and actual data on the amount of power generated by wind power generation in a wide area. Here, as with FIG. 2, an example will be described in which January 2nd is the target prediction date. The performance evaluation unit 15 can generate each graph shown in FIG. 3, for example. For example, the performance evaluation unit 15 generates a graph using graph 121 of the forecast results and graph 131 of the actual results as the results of performance evaluation for December 31st. Furthermore, the performance evaluation unit 15 generates a graph using graph 122 of the forecast results and graph 132 of the actual results before 9:00 as the results of performance evaluation for January 1st. Furthermore, the performance evaluation unit 15 generates a graph using graph 123 of the forecast results for January 2nd.

[0068] Continuing the explanation, returning to Figure 1, the result output unit 18 receives an input of the prediction result of the amount of power generated by wind power generation in the wide area on the prediction target day from the prediction execution unit 14. The result output unit 18 also receives an input of the result of performance evaluation of the day before the prediction target day and the performance data of the amount of power generated by wind power generation in the wide area acquired up to that point from the performance evaluation unit 15.

[0069] The result output unit 18 then stores the obtained prediction results of the amount of wind power generated in the wide area on the target prediction day in folders for the target prediction day, the day before the target prediction day, and the day before the target prediction day, respectively. That is, the prediction results for each day up to two days after the target prediction day are ultimately stored in the folder for each day. At this time, the result output unit 18 assigns a creation date to the file indicating each piece of information. This allows the result output unit 18 to record information on the timing of the prediction and performance evaluation. The result output unit 18 may also store the performance data and performance evaluation results for that day in the folder for each day.

[0070] The result output unit 18 then displays the prediction results, performance data, and performance evaluation results on a monitor or the like to provide the user with information. For example, the result output unit 18 provides the user with graphs and tables showing the prediction results and performance data, respectively. This allows the user to visually confirm the difference between the prediction results and performance data, which can be used to determine the reliability of the prediction. Alternatively, the result output unit 18 may provide the user with information by sending the prediction results, performance data, and performance evaluation results to a designated external terminal device (not shown) or the like.

[0071] The result output unit 18 may also store folders for each day containing the predicted results, performance data, and performance evaluation results in an external server (not shown). For example, the user may access the server using their own terminal device to obtain the data. If the server has a mail server function, the user may periodically send data to the mail server to provide the information to the user.

[0072] When the prediction execution unit 14 performs prediction results after 9:00 on January 1st, the result output unit 18 stores the prediction results in folders for December 31st, January 1st, and January 2nd, respectively. In this case, the December 31st folder stores the prediction results for each of December 31st, January 1st, and January 2nd, the performance data for December 31st, and the performance evaluation result for December 31st. Furthermore, the January 1st folder stores the prediction results for each of January 1st and January 2nd, and the performance data up to 9:00 on January 1st. Furthermore, the January 2nd folder stores the prediction results for January 2nd.

[0073] 9 is a flowchart of the machine learning process performed by the wind power generation prediction device according to Example 1. Next, the flow of the machine learning process in the learning phase performed by the wind power generation prediction device 1 according to Example 1 will be described with reference to FIG.

[0074] The information collection unit 11 executes scraping of various data used to predict the amount of power generated by the wind power generator to be used for training the machine learning model 17 (step S1). Then, the information collection unit 11 stores the collected data in the data storage unit 12.

[0075] The aggregation and consolidation unit 13 performs preprocessing including aggregation and consolidation on the data stored in the data storage unit 12 (step S2).

[0076] The learning execution unit 16 receives input of learning data including actual data on the amount of power generated by wind power generation in a wide area and predicted data on wind speed from the aggregation and organization unit 13. Next, the learning execution unit 16 standardizes the predicted data on wind speed (step S3).

[0077] Next, the learning execution unit 16 sets the specified learning conditions in the machine learning model 17 (step S4).

[0078] Next, the learning execution unit 16 trains the machine learning model 17 by mini-batch learning to adjust the parameters (step S5). Specifically, the learning execution unit 16 inputs the predicted data of wind speed every 30 minutes for the 24 hours before 9:00 on the prediction target date to the encoder 171 of the machine learning model 17. The learning execution unit 16 also inputs the predicted data of wind speed every 30 minutes from 9:00 on the day before the prediction target date until midnight on the prediction target date, which is 39 hours later, to the decoder 172 of the machine learning model 17. The learning execution unit 16 then compares the predicted values of the amount of power every 30 minutes from 9:00 on the day before the prediction target date until midnight on the prediction target date, which is 39 hours later, output from the machine learning model 17, with the corresponding actual data, and causes the machine learning model 17 to perform learning.

[0079] Next, the learning execution unit 16 checks whether overlearning has occurred (step S6).

[0080] The learning execution unit 16 determines whether or not overlearning has occurred based on the confirmed result (step S7). If overlearning has occurred (step S7: Yes), the learning execution unit 16 returns to step S6 and causes the machine learning model 17 to perform learning again so as to eliminate the overlearning.

[0081] On the other hand, if overlearning has not occurred (step S7: No), the learning execution unit 16 obtains a trained machine learning model 17 and completes model creation (step S8).

[0082] 10 is a flowchart of the power generation amount prediction process performed by the wind power generation prediction device according to the embodiment 1. Next, the flow of the power generation amount prediction process in the prediction phase performed by the wind power generation prediction device 1 according to the embodiment 1 will be described with reference to FIG.

[0083] The information collection unit 11 executes scraping of various data used for learning the machine learning model 17 or for predicting the amount of power generated by wind power generators in a predetermined area (step S11).

[0084] The aggregation and consolidation unit 13 performs pre-processing including aggregation and consolidation on the data stored in the data storage unit 12 (step S12).

[0085] The prediction execution unit 14 receives wind speed forecast data from the aggregation and organization unit 13. Next, the prediction execution unit 14 performs standardization processing on the acquired wind speed forecast data. The prediction execution unit 14 then inputs the processed wind speed forecast data into the trained machine learning model 17 to obtain a forecast value for the amount of power generated by wind power generation in the wide area. In this way, the prediction execution unit 14 executes a forecast of the amount of power generated by wind power generation in the wide area (step S13). Specifically, the prediction execution unit 14 inputs wind speed forecast data for every 30 minutes for the 24 hours before 9:00 a.m. on the day before the target prediction day to the encoder 171 of the machine learning model 17. The prediction execution unit 14 also inputs wind speed forecast data for every 30 minutes from 9:00 a.m. on the day before the target prediction day until midnight on the target prediction day, which is 39 hours after the target prediction day, to the decoder 172 of the machine learning model 17. The learning execution unit 16 then obtains the amount of power for every 30 minutes up to midnight on the prediction target day, which corresponds to 39 hours from 9:00 on the day before the measurement target day, output from the machine learning model 17. The prediction execution unit 14 then stores the prediction result of the amount of power generated by wind power generation in the wide area on the prediction target day in the data storage unit 12.

[0086] The performance evaluation unit 15 acquires from the data storage unit 12 the forecast result of the amount of power generated by wind power generation in the wide area on the day before the target prediction date and the performance data of the amount of power generated by wind power generation in the wide area on the day before the target prediction date. Next, the performance evaluation unit 15 compares the forecast result of the amount of power generated by wind power generation in the wide area on the day before the target prediction date with the performance data of the amount of power generated by wind power generation in the wide area on the day before the target prediction date to evaluate the performance. For example, the performance evaluation unit 15 generates a graph showing the forecast value and the actual value of the amount of power generated by wind power generation in the wide area on the day before the target prediction date, and a graph showing the forecast value and the actual value of the amount of power generated by wind power generation in the wide area on the day before the target prediction date up to the reference time. The performance evaluation unit 15 also generates a graph showing the forecast value of the amount of power generated by wind power generation in the wide area on the target prediction date. Furthermore, the performance evaluation unit 15 creates a table showing the results of the performance evaluation (step S14).

[0087] The result output unit 18 acquires the results of the performance evaluation by the performance evaluation unit 15, i.e., the graphs and tables created by the performance evaluation unit 15. Then, the result output unit 18 notifies the results by displaying the results of the performance evaluation including the acquired graphs and tables on a monitor or the like (step S15).

[0088] FIG. 11 is a diagram illustrating a comparison of prediction accuracy. In FIG. 11, the column labeled LSTM represents the evaluation results of the prediction accuracy by the wind power generation prediction device 1 according to the first embodiment. Furthermore, multiple regression, RFR (Random Forest Regression), and SVR (Support Vector Regression) are used as methods for comparison, and the columns corresponding to the names of each represent the evaluation results of the prediction accuracy. Here, RMSE (Root Mean Squared Error) was used as the evaluation method for prediction accuracy. In the evaluation results shown in FIG. 11, a smaller value indicates better accuracy.

[0089] 11, the prediction by the wind power generation prediction device 1 according to the first embodiment shows the best and second best daily accuracy on many days. Also, looking at the overall picture, the prediction by the wind power generation prediction device 1 according to the first embodiment shows the best accuracy. In other words, it can be said that the prediction by the wind power generation prediction device 1 according to the first embodiment is more accurate than when other prediction methods are used.

[0090] Fig. 12 is a diagram showing the daily transition of the predicted results of power generation using a plurality of prediction methods and the actual values of power generation. In Fig. 8, the horizontal axis represents the number of frames, with each frame being 30 minutes, and the vertical axis represents the power generation amount of wind power in a wide area. Graph 211 shows the actual value of power generation amount of wind power in a wide area on September 9th. Graph 212 shows the prediction results for September 9th by the wind power generation prediction device 1 according to Example 1. Graph 213 shows the prediction results by multiple regression prediction, graph 214 shows the prediction results by RFR prediction, and graph 215 shows the prediction results by SVR prediction.

[0091] 12, it can be said that the graph 212 showing the prediction results by the wind power generation prediction device 1 according to the first embodiment closely follows the changes in the graph 211 showing the actual values, compared to the graphs 213 to 215 showing the prediction results when other prediction methods are used. In this way, the predictions by the wind power generation prediction device 1 according to the first embodiment tend to fit best to the actual values.

[0092] As described above, the wind power generation prediction device 1 according to the first embodiment causes the machine learning model 17, which uses a sequence-to-sequence model, to perform learning by mini-batch learning using the wind speed at each point in a wide area and the actual values of the power generation amount of wind power generation in the wide area. The wind power generation prediction device 1 also grasps the overlearning situation and generates a trained machine learning model 17. Then, the wind power generation prediction device 1 predicts the power generation amount of wind power generation in the wide area from the predicted data of the wind speed at each point in the wide area using the trained machine learning model 17.

[0093] In this way, the wind power generation prediction device 1 can predict the amount of wind power generation over a wide area, rather than predicting the amount of power generation for each individual wind power generator or for each site of a group of wind power generators. Furthermore, since predictions are made without using detailed information about each wind power generator, it is not necessary to acquire a prediction model that converts a predicted wind speed value using a power curve to calculate a predicted power output value. Furthermore, by using a sequence-to-sequence model such as LSTM as the machine learning model 17, predictions can be made based on time-series data, making it possible to more accurately predict the amount of wind power generation that changes over time. Furthermore, advanced learning can be performed using mini-batch learning, improving prediction accuracy. Furthermore, by generating the machine learning model 17 taking overlearning into consideration, it is possible to achieve predictions with higher prediction accuracy. Therefore, the wind power generation prediction device 1 according to the first embodiment can efficiently and accurately predict the amount of power generation from wind power generators.

[0094] For example, there are multiple new power balancing groups in the Tohoku region. The supply and demand structure within each balancing group changes frequently due to contract changes and other factors. Therefore, each balancing group must forecast power generation to submit a next-day supply and demand plan, and appropriate supply is required based on the current supply and demand structure. However, some balancing groups are small and do not have conventional power generation forecasting technology. In particular, after the end of the feed-in tariff system, it is expected that the number of businesses seeking forecasting technology will increase. Therefore, even in balancing groups that include businesses without forecasting technology, the wind power generation forecasting device 1 according to the first embodiment can accurately forecast wind power generation, thereby optimizing the amount of power purchased from the market and enabling appropriate next-day supply and demand balance plans. Furthermore, from the perspective of power producers, if they can accurately forecast next-day power generation, they can trade with the Japan Electric Power Exchange (JEPX) at market prices after the end of the feed-in tariff system, rather than trading with balancing groups such as electric power companies at surplus power generation prices. This increased choice enables more efficient trading.

[0095] (Variation) Here, the wind power generation prediction device 1 according to the first embodiment uses prediction data for wind speeds at times that have already passed, but it is also possible to use actual measurement data of wind speeds for past time periods, for example. The operation of each unit in this case will be described below. The description of the operation of each unit that is the same as that of the first embodiment will be omitted.

[0096] The information collection unit 11 collects wind speed measurement data for the wide area before the reference time on the target prediction day. For example, the information collection unit 11 collects wind speed measurement data for each hour from 9:00 to 24 hours before.

[0097] The learning execution unit 16 performs standardization processing on the measured wind speed data for the period from 9:00 to 24 hours before, and then causes the machine learning model 17 to perform learning.

[0098] The prediction execution unit 14 standardizes the measured wind speed data for the period from 9:00 to 24 hours before, and then inputs the data into the machine learning model 17 to obtain a prediction result.

[0099] In this way, by using actual measured wind speed data for past data that can be obtained, it is possible to further improve the prediction accuracy. [Example]

[0100] Next, a second embodiment will be described. A wind power generation prediction device 1 according to the second embodiment is also shown in the block diagram of FIG. 1. The wind power generation prediction device 1 according to the second embodiment predicts the amount of wind power generation in a wide area using the wind direction at each point in the wide area as well as the wind speed. Details of the wind power generation prediction device 1 according to the second embodiment will be described below. In the following description, explanations of the operations of the same parts as those in the first embodiment will be omitted.

[0101] In addition to the wind speed forecast data, the information collection unit 11 collects hourly wind direction forecast data for each location in the wide area from 24 hours before the reference time until midnight on the forecast target day from the weather information system 2. The information collection unit 11 then stores the wind direction forecast data for each location in the wide area on the forecast target day together with the wind speed forecast data in the data storage unit 12. At this time, the information collection unit 11 overwrites the forecast data while complementing missing data, just as with the wind speed.

[0102] The machine learning model 17 according to the second embodiment is a prediction model that predicts the amount of power generated by wind power generation in a wide area using predicted data of wind speed and wind direction as input.

[0103] 13 is a diagram illustrating a wind power generation prediction algorithm according to Example 2. The machine learning model 17 according to Example 2 also includes an encoder 171 that is an input-side LSTM and a decoder 172 that is an output-side LSTM.

[0104] In FIG. 13, wind direction ##1 represents the forecast data for wind direction at 9:00 a.m. on the day before the forecast target date for 33 locations in a wide area. Wind directions ##2 to ##79 represent the forecast data for wind direction every 30 minutes for the 39 hours after 9:00 a.m. on the day before the forecast target date. Wind directions ##0 to ##-47 represent the forecast data for wind direction every 30 minutes within the 24 hours before 9:00 a.m. Here, the machine learning model 17 uses quantified wind direction forecast data as described below.

[0105] The encoder 171 has 48 combinations of a coupled layer and the LSTM encoders 710 connected to it, and each LSTM encoder 710 is connected in series. For each combination, the past 24 hours of actual data (wind speed #-47, wind direction ##-47) to (wind speed #0, wind direction ##0) are input to the coupled layer in the order they are arranged in series.

[0106] Fig. 14 is a diagram showing an example of the configuration of the combined layer. For example, a case where wind speed #t and wind direction ##t are input to the combined layer will be described. For example, wind speed #t is expressed as a vector having standardized values of wind speed at each of 33 locations. Furthermore, wind direction ##t is expressed as a numerical value between 0 and 16 through embedding processing.

[0107] The wind speed #t, which is a continuous value, is input to the fully connected layer and reaches the hidden layer 731. The quantified wind direction ##t, which is a discrete value, is converted to a hidden layer 732 of the same dimension as the hidden layer 731 by embedding. An adder 733 then performs a calculation on the output from the hidden layer 731 and the output from the hidden layer 732. The adder 733 calculates the element-by-element sum of the hidden layer 731 and the hidden layer 732. Next, an accumulator 734 performs a calculation on the output from the hidden layer 731 and the output from the adder 733. The accumulator 734 calculates the element-by-element product of the output from the hidden layer 731 and the output from the adder 733. The output from the accumulator 734 is then input to the LSTM encoder 710 as the output from the connected layer. By using such a connected layer to combine wind speed information and wind direction information, the accuracy of forecasting the amount of wind power generation over a wide area can be improved.

[0108] Here, the value output from hidden layer 731, obtained by inputting the standardized wind speed prediction data into machine learning model 17, is an example of a "first value." Furthermore, the value output from hidden layer 732, obtained by inputting the digitized wind direction prediction data into machine learning model 17, is an example of a "second value." The second value is obtained by performing a process on the wind direction prediction data to match the dimension of the first value. Then, machine learning model 17 adds the first value output from hidden layer 731 and the second value output from hidden layer 732, and predicts the power generation amount of wind power generators in a wide area based on the value obtained by multiplying the sum by the first value.

[0109] Returning to Figure 13, the explanation continues. Predicted data for the corresponding wind speed and wind direction is input to each LSTM encoder 710 via the coupling layer. Each LSTM encoder 710 other than the first in the series is input with the information output from the previous LSTM encoder 710. The last LSTM encoder 710 in the series then outputs the latent feature 173.

[0110] In this way, the LSTM encoder 710 can obtain latent features 173 that incorporate all of the time-series data, wind speeds #-47 to #0 and wind directions ##-47 to ##0, by sequentially propagating the outputs obtained from the wind speed and wind direction at each time. Conceptually, these latent features 173 can be said to be information summarizing the wind speed and wind speed conditions over the most recent 24 hours.

[0111] The decoder 172 has 79 combinations of a front-stage coupled layer, an LSTM decoder 720, and a rear-stage fully coupled layer, and each LSTM decoder 720 is connected in series. For each combination, (wind speed #1, wind direction ##1) to (wind speed #79, wind direction ##79) obtained as prediction data for 39 hours from 9:00 are input to the front-stage coupled layer in the order they are arranged in series. The front-stage coupled layer also has, for example, the configuration shown in FIG. 14.

[0112] Each LSTM decoder 720 receives input of the corresponding predicted wind speed and direction data via the previous connected layer. The first LSTM decoder 720 in the series receives input of the latent feature 173 output from the encoder 171. Each LSTM decoder 720 other than the first receives input of the information output from the previous LSTM decoder 720. The subsequent fully connected layer receives input of the value output from the LSTM decoder 720, and outputs power generation E1 to E79, which are predicted values of wind power generation in a wide area including the corresponding time of the target prediction day.

[0113] In this way, each LSTM decoder 720 can make a prediction that takes into account the wind speed and wind direction from the previous 24 hours up to midnight on the prediction target day by sequentially propagating the wind speed and wind direction information for the previous 24 hours. Therefore, the decoder 172 can take into account the most recent wind speed and wind direction states in chronological order, as well as the wind speed and wind direction on the prediction target day in chronological order, when predicting the amount of power generation, improving the accuracy of the prediction of wind power generation over a wide area in chronological order.

[0114] In the learning phase, the learning execution unit 16 receives input of learning data including forecast data of wind speed and wind direction and actual data of power generation amount of wind power generation in a wide area from the aggregation and organization unit 13. The learning execution unit 16 standardizes the forecast data of wind speed.

[0115] The learning execution unit 16 also performs a digitization process on the predicted wind direction data to represent the wind direction numerically. For example, the learning execution unit 16 digitizes each of the 16 directions by assigning numbers from 1 to 15 clockwise, starting with north as 0. The learning execution unit 16 also interpolates the hourly predicted wind direction data to convert it into predicted data for every 30 minutes.

[0116] The learning execution unit 16 then inputs the standardized wind speed prediction data and the digitized wind speed prediction data to the machine learning model 17. For example, the learning execution unit 16 inputs the wind speed and wind direction prediction data for every 30 minutes within the 24 hours before 9:00 on the day before the target prediction day to the encoder 171 of the machine learning model 17. The learning execution unit 16 also inputs the wind speed and wind direction prediction data for every 30 minutes for 39 hours from 9:00 on the day before the target prediction day until midnight on the target prediction day to the decoder 172 of the machine learning model 17.

[0117] Then, the learning execution unit 16 divides the learning data into multiple groups and causes the machine learning model 17 to perform learning using mini-batch learning. After that, if overfitting occurs, the learning execution unit 16 performs learning again to generate a trained machine learning model 17 in which overfitting has been suppressed.

[0118] The prediction execution unit 14 receives prediction data including predicted data of wind speed and predicted data of wind direction from the aggregation and organization unit 13. Similar to the learning execution unit 16, the prediction execution unit 14 performs standardization processing of the predicted data of wind speed, and digitization processing and interpolation processing of the predicted data of wind direction.

[0119] Next, the prediction execution unit 14 inputs the standardized wind speed prediction data and the quantified wind direction prediction data into the machine learning model 17. Thereafter, the prediction execution unit 14 acquires the predicted values of the amount of power generated by wind power generation in the wide area for every 30 minutes from the reference time on the day before the prediction target day until midnight on the prediction target day, which are output from the decoder 172. The prediction execution unit 14 then extracts the predicted values for every 30 minutes for the 24 hours on the prediction target day from the acquired predicted values, and sets these as the prediction results of the amount of power generated by wind power generation in the wide area on the prediction target day.

[0120] Here, the wind power generation prediction device 1 according to the second embodiment uses prediction data for the wind direction of a time that has already passed, but it is also possible to use actual measurement data of the wind direction for a time period in the past, similar to the wind speed.

[0121] As described above, the wind power generation prediction device 1 according to the second embodiment trains the machine learning model 17 using the prediction data of wind speed and wind direction, which are time-series data. Then, the wind power generation prediction device 1 inputs the prediction data of wind speed and wind direction, which are time-series data, into the trained machine learning model 17, and predicts the amount of power generated by wind power generation on the prediction target day in a wide area, which is time-series data. In this way, by predicting the amount of power generated by wind power generation in a wide area using the machine learning model 17 that has been trained taking into account wind direction in addition to wind speed, it is possible to make a prediction that takes into account more information indicating wind conditions, and it is possible to improve the prediction accuracy. [Example]

[0122] Next, a third embodiment will be described. A wind power generation prediction device 1 according to the third embodiment is also shown in the block diagram of FIG. 1. In the learning phase, the wind power generation prediction device 1 according to the third embodiment causes a machine learning model 17 to perform learning by adding a weight according to the accuracy of wind speed prediction data. Details of the wind power generation prediction device 1 according to the third embodiment will be described below. In the following description, a description of the operation of each unit similar to that of the first or second embodiment will be omitted.

[0123] The information collection unit 11 collects actual wind speed measurement data for the wide area before the reference time on the prediction target day in addition to the predicted wind speed and wind direction data. The information collection unit 11 then stores the actual wind speed measurement data in the data storage unit 12.

[0124] In addition to the predicted data of wind speed and wind direction, the aggregation and organization unit 13 acquires actual measured data of wind speed for the same time from the data storage unit 12, performs preprocessing on the data together with the wind speed and wind direction, and outputs the data to the learning execution unit 16. In this case, for a time when actual measured data is missing, the aggregation and organization unit 13 may exclude the predicted data of wind speed and wind direction for that time from the learning data.

[0125] The learning execution unit 16 stores in advance the following mathematical formula (3) which indicates the weight at time t: tis the root mean square error between wind speed #t, which is the predicted value of wind speed at time t, and the actual measured value. Furthermore, T is the last time of the prediction target, which represents 24:00 on the prediction target day in Example 3. This weight is heavier the smaller the difference between the predicted value and the actual value. In other words, by using this weight, it becomes possible to have the machine learning model 17 perform learning by placing more importance on wind speed prediction data when the difference between the predicted value and the actual value is small.

[0126]

number

[0127] The learning execution unit 16 holds the following formula (4) as a formula indicating the loss in the prediction of the amount of power generation at time t, which is indicated using the weight indicated by formula (3). w t × ((Predicted power generation amount at time t) - (Actual power generation amount at time t)) 2 ···(4)

[0128] The learning execution unit 16 uses the acquired predicted data and measured data of wind speed for each time in equation (3) to calculate the weight for each time from the reference time to midnight on the prediction target day. Next, the learning execution unit 16 uses the weight for each time, the predicted value of power generation amount, and the actual value obtained from the actual data of power generation amount in equation (4) to calculate the loss in the prediction of power generation amount for each time and generate a loss function. Then, the learning execution unit 16 causes the machine learning model 17 to perform learning so as to minimize the generated loss function.

[0129] The learning execution unit 16 causes the machine learning model 17 to learn by mini-batch learning using the predicted data of wind speed, predicted data of wind direction, and the generated loss function. Here, in the third embodiment, the case where both wind speed and wind direction are used has been described, but the same applies to the case where the machine learning model 17 is caused to learn using predicted data of wind speed without using predicted data of wind direction.

[0130] In this way, the learning execution unit 16 calculates the weights for each piece of predicted wind speed data based on the predicted wind speed data and the actual measured wind speed data, and causes the machine learning model to learn using the calculated weights, the predicted wind speed data, and the actual power generation data as learning data.

[0131] 15 is a flowchart of the machine learning process performed by the wind power generation prediction device according to Example 3. Next, the flow of the machine learning process performed by the wind power generation prediction device 1 according to Example 3 will be described with reference to FIG.

[0132] The information collection unit 11 executes scraping of various data used to predict the amount of power generated by the wind power generator, which is used for training the machine learning model 17, including predicted data of wind speed and predicted data of wind direction (step S101). Then, the information collection unit 11 stores the collected data in the data storage unit 12.

[0133] The aggregation and consolidation unit 13 performs pre-processing including aggregation and consolidation on the data stored in the data storage unit 12 (step S102).

[0134] The learning execution unit 16 receives input of learning data including actual data on the amount of power generated by wind power generation in a wide area and predicted data on wind speed from the aggregation and organization unit 13. Next, the learning execution unit 16 standardizes the predicted data on wind speed (step S103).

[0135] Next, the learning execution unit 16 performs averaging on the standardized predicted wind speed data to obtain predicted wind speed data for every 30 minutes (step S104).

[0136] Next, the learning execution unit 16 performs a digitization process on the predicted wind direction data (step S105). Furthermore, the learning execution unit 16 interpolates the digitized predicted wind direction data to obtain predicted wind speed data for every 30 minutes.

[0137] Next, the learning execution unit 16 sets the specified learning conditions in the machine learning model 17 (step S106).

[0138] The learning execution unit 16 inputs the predicted wind speed data and the predicted wind direction data into the machine learning model 17, and causes the model 17 to execute prediction of the amount of power generated by the wind power generators in a wide area, thereby obtaining a predicted value of the amount of power generated (step S107).

[0139] Next, the learning execution unit 16 calculates the weight for each time period from the reference time to midnight on the target prediction day from the difference between the predicted value obtained from the predicted wind speed data and the actual measured value obtained from the actual measured wind speed data (step S108).

[0140] Next, the learning execution unit 16 calculates the loss in the prediction of the amount of power generation for each time period using the weight, the predicted value of the amount of power generation, and the actual value obtained from the actual data of the amount of power generation, and generates a loss function (step S109).

[0141] Next, the learning execution unit 16 adjusts the parameters of the machine learning model 17 so as to reduce the loss function (step S110). The learning execution unit 16 causes the machine learning model 17 to execute the learning from steps S107 to S110 by mini-batch learning.

[0142] The learning execution unit 16 determines whether a predetermined learning termination condition has been reached (step S111). If the termination condition has not been reached (step S111: No), the learning execution unit 16 returns to step S107 and repeats learning by mini-batch learning.

[0143] On the other hand, if the termination condition is reached (step S111: Yes), the learning execution unit 16 checks whether the learning is over-learned (step S112).

[0144] The learning execution unit 16 determines whether or not overlearning has occurred based on the confirmed result (step S113). If overlearning has occurred (step S113: Yes), the learning execution unit 16 returns to step S107 and causes the machine learning model 17 to perform learning again so as to eliminate the overlearning.

[0145] On the other hand, if overlearning has not occurred (step S113: No), the learning execution unit 16 obtains a trained machine learning model 17 and completes model creation (step S114).

[0146] As described above, the wind power generation prediction device 1 according to the third embodiment generates a loss function by adding a weight to the wind speed for each hour of the day, and causes the machine learning model 17 to learn the loss function. By using the weight, it becomes possible to cause the machine learning model 17 to perform learning by placing more importance on the predicted data of the wind speed when the difference between the predicted value and the actual value is small, and it becomes possible to further improve the prediction accuracy. Furthermore, although the above description has been given of an example in which a function for calculating a loss function using a weight is added to the second embodiment, it is also possible to add this function to the first embodiment, and the above effects can be obtained.

[0147] (Hardware configuration) Fig. 16 is a hardware configuration diagram of the wind power generation prediction device. Here, an example of a hardware configuration for realizing each function of the wind power generation prediction device 1 according to each embodiment and modification will be described with reference to Fig. 16.

[0148] 16, the wind power generation prediction device 1 according to each embodiment includes a CPU (Central Processing Unit) 91, a memory 92, a hard disk 93, and a network interface 94. The CPU 91 is connected to the memory 92, the hard disk 93, and the network interface 94 via a bus.

[0149] The network interface 94 is an interface for communication between the wind power generation prediction device 1 and an external device. The network interface 94 relays communication between the CPU 91 and the weather information system 2 or the wind power generator power generation record database 3, for example.

[0150] The hard disk 93 is an auxiliary storage device. The hard disk 93 realizes the function of the data storage unit 12. The hard disk 93 may also store a machine learning model 17. The hard disk 93 also stores various programs including programs for realizing the functions of the information collection unit 11, the aggregation and organization unit 13, the prediction execution unit 14, the performance evaluation unit 15, the learning execution unit 16, and the result output unit 18 illustrated in FIG. 1.

[0151] The memory 92 is a main storage device and may be, for example, a dynamic random access memory (DRAM).

[0152] The CPU 91 reads various programs from the hard disk 93, expands them into the memory 92, and executes them. As a result, the CPU 91 realizes the functions of the information collection unit 11, the aggregation and organization unit 13, the prediction execution unit 14, the performance evaluation unit 15, the learning execution unit 16, and the result output unit 18 illustrated in FIG. [Explanation of symbols]

[0153] 1. Wind power generation forecasting device 2. Weather Information System 3. Wind power generation performance database 11 Information Gathering Department 12 Data storage section 13 Consolidation and Arrangement Department 14 Prediction Execution Department 15 Performance Evaluation Department 16 Learning Execution Department 17 Machine Learning Models 18 Result output section

Claims

1. a machine learning model configured by serially connecting time series transformation models between time series data; an information acquisition unit that acquires forecast data of wind speeds at multiple points in a predetermined area where multiple wind power generators are installed in various locations and actual data of the total power generation amount of the multiple wind power generators at the time of prediction of the forecast data of the wind speeds; a learning execution unit that causes the machine learning model to learn the predicted data of the wind speed and the actual data of the total power generation amount as learning data; a prediction execution unit that inputs the predicted data of the wind speed into the trained machine learning model and predicts the total power generation amount of the plurality of wind power generators; A wind power generation prediction device comprising:

2. the information acquisition unit further acquires forecast data of wind directions at the plurality of points; the learning execution unit causes the machine learning model to learn using the predicted data of the wind speed, the predicted data of the wind direction, and the actual data of the total power generation amount as learning data; The prediction execution unit inputs the wind speed prediction data and the wind direction prediction data into the machine learning model to predict the total power generation amount of the plurality of wind power generators. The wind power generation prediction device according to claim 1 .

3. The wind power generation prediction device according to claim 2, characterized in that the machine learning model obtains a first value from the wind speed prediction data, performs a process for aligning the dimensions of the wind direction prediction data with the first value to obtain a second value, adds the first value and the second value, and predicts the total power generation amount of the wind power generators based on the value obtained by multiplying the sum by the first value.

4. the information acquisition unit further acquires actual measurement data of wind speeds at the plurality of points at a time point when the wind speed prediction data is predicted; The learning execution unit calculates weights for the predicted wind speed data based on the predicted wind speed data and the actual measured wind speed data, and causes the machine learning model to learn using the calculated weights, the predicted wind speed data, and the actual data on the total amount of power generation as learning data. The wind power generation prediction device according to claim 1 .

5. A wind power generation forecasting device having a machine learning model configured by serially connecting series transformation models between series data, Obtaining wind speed forecast data for a plurality of points in a predetermined area where a plurality of wind power generators are installed in various locations and actual data on the total amount of power generated by the plurality of wind power generators at the time of prediction of the wind speed forecast data; the predicted data of the wind speed and the actual data of the total power generation amount are used as learning data to train the machine learning model; The predicted data of the wind speed is input into the trained machine learning model to predict the total power generation amount of the plurality of wind power generators. A wind power generation forecasting method comprising the steps of:

6. Obtaining wind speed forecast data for a plurality of points in a predetermined area where a plurality of wind power generators are installed in various locations and actual data on the total amount of power generated by the plurality of wind power generators at the time of prediction of the wind speed forecast data; the predicted data of the wind speed and the actual data of the total power generation amount are used as learning data, and a machine learning model having a configuration in which time-series data series transformation models are connected in series is trained; The predicted data of the wind speed is input into the trained machine learning model to predict the total power generation amount of the plurality of wind power generators. A wind power generation forecasting program that causes a computer to execute processing.

Images