Wind power generation forecasting device and wind power generation forecasting method

Abstract

This invention provides a wind power generation forecasting device and a wind power generation forecasting method that can improve the accuracy of wind power generation forecasts by accurately predicting wind speed. [Solution] The system includes a data processing unit 7 that processes weather observation data from multiple wind turbines 17 of a wind power generation facility and weather forecast data including representative values ​​of ensemble weather forecast data acquired from an external source for the purpose of learning and inferring a wind speed prediction model; a wind speed prediction model learning unit 9 that generates a wind speed prediction model that predicts the wind speed at the installation sites of the multiple wind turbines 17 using the data processed by the data processing unit 7; and a power generation amount calculation unit 11 that infers and calculates the wind speed at the installation sites of the multiple wind turbines 17 using the wind speed prediction model and calculates the amount of power generated by the multiple wind turbines 17 based on a power curve.

Description

[Technical Field]

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

[0002] Wind power generation utilizes the mechanism of a wind turbine to generate electricity, but the amount of electricity generated depends heavily on the wind speed at the location where the turbine is installed. Therefore, accurately predicting the wind speed at the turbine installation site is necessary to forecast the amount of electricity generated. In particular, in order to conduct bidding for the buying and selling of generated electricity in the electricity market, it is necessary to forecast the wind speed for a predetermined period and then make a highly accurate forecast of the amount of electricity generated from that forecast. In other words, medium- and long-term forecasts of the amount of electricity generated for power generation projects must be based on wind speed forecasts.

[0003] Patent Document 1 discloses a method for predicting the wind speed distribution of an evaluation area from meteorological data and actual wind condition measurement data. A neural network is used for the prediction. Patent Document 2 discloses a method for calculating predicted values ​​for power generation at a specified date and time based on a prediction formula using multiple regression analysis that shows the relationship between the power generation amount of a power source and meteorological information of the power source's location. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2023-7932 [Patent Document 2] Japanese Patent Publication No. 2016-73156 [Non-patent literature]

[0005] [Non-Patent Document 1] Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O. Arik, Tomas Pfister, “TSMixer: An All-MLP Architecture for Time Series Forecasting”, [online], Transactions on Machine Learning Research (TMLR), 09 / 2023, [Retrieved February 14, 2025], Internet,<URL:https: / / arxiv.org / abs / 2303.06053> [Overview of the project] [Problems that the invention aims to solve]

[0006] In Patent Document 1, there are doubts as to whether appropriate meteorological data and wind condition measurement data are set as inputs for high-precision prediction. In Patent Document 2, multiple regression analysis is used to calculate multiple prediction patterns, prediction reliability, and estimated total power generation, which may lead to high computational costs.

[0007] This invention has been made in view of the circumstances described above, and the problem that this invention aims to solve is to provide a wind power generation prediction device and a wind power generation prediction method that can improve the accuracy of wind power generation prediction by predicting wind speed with high accuracy. [Means for solving the problem]

[0008] To solve the above problems, the present invention employs the following means. The wind power generation prediction device of the present invention comprises: a data processing unit that processes weather observation data from multiple wind turbines for wind power generation and weather forecast data including representative values ​​of ensemble weather forecast data acquired from an external source for the purpose of learning and inferring a wind speed prediction model; a wind speed prediction model learning unit that generates a wind speed prediction model that predicts the wind speed at the installation sites of the multiple wind turbines using the data processed by the data processing unit; and a power generation calculation unit that infers and calculates the wind speed at the installation sites of the multiple wind turbines using the wind speed prediction model and calculates the amount of power generated by the multiple wind turbines based on a power curve.

[0009] In a wind power generation forecasting device according to one aspect of the present invention, the representative value of the ensemble weather forecast data is the statistical value of the ensemble weather data or the value of some or all members of the ensemble weather data.

[0010] In a wind power generation prediction device according to one aspect of the present invention, the learning of the wind speed prediction model uses a function that evaluates the reliability of weather forecast data, in addition to the mean squared error function between the training value and the predicted value, as the loss function.

[0011] The wind power generation prediction method of the present invention includes processing meteorological observation data from multiple wind turbines and meteorological forecast data including representative values ​​of ensemble meteorological forecast data obtained from an external source for the purpose of training and inference of a wind speed prediction model; generating a wind speed prediction model that predicts the wind speed at the installation sites of the multiple wind turbines using the processed data; inferring and calculating the wind speed at the installation sites of the multiple wind turbines using the wind speed prediction model, and calculating the amount of power generated by the multiple wind turbines based on a power curve.

[0012] In a wind power generation forecasting method according to one aspect of the present invention, the representative value of the ensemble weather forecast data is the statistical value of the ensemble weather data or the value of some or all members of the ensemble weather data.

[0013] In the wind power generation amount prediction method according to one aspect of the present invention, in the training of the wind speed prediction model, a function for evaluating the reliability of weather forecast data is used as a loss function in addition to the mean square error function of a teacher value and a predicted value. Effects of the Invention

[0014] According to the present invention, it is possible to provide a wind power generation amount prediction apparatus and a wind power generation amount prediction method that can improve the accuracy of wind power generation amount prediction by accurately predicting wind speed. Brief Description of the Drawings

[0015] [Figure 1] It is a block diagram showing the configuration of a wind power generation amount prediction apparatus according to an embodiment of the present invention. [Figure 2] It is a schematic diagram showing a configuration example of a prediction model of a wind power generation amount prediction apparatus according to an embodiment of the present invention. Mode for Carrying Out the Invention

[0016] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. Embodiment FIG. 1 is a block diagram showing the configuration of a wind power generation amount prediction apparatus 1 according to an embodiment of the present invention. As shown in FIG. 1, the wind power generation amount prediction apparatus 1 includes a reception unit 3, a storage unit 5, a data processing unit 7, a wind speed prediction model learning unit 9, a power generation amount calculation unit 11, and an output unit 15. Outside the wind power generation amount prediction apparatus 1, a weather prediction apparatus 13 and a plurality of wind turbines 17 for wind power generation are installed, and the reception unit 3 receives various data transmitted from the weather prediction apparatus 13 and the plurality of wind turbines 17.

[0017] The weather forecasting device 13 receives weather forecast data from an external source or creates weather forecast data using its own weather forecasting model, and transmits that data to the receiving unit 3. Each of the wind turbines 17 is equipped with a weather observation device 17a, a power generation measurement device 17b, and a wind turbine operation log recording device 17c. The weather observation device 17a acquires weather observation data (wind speed, wind direction, etc.) at the location where the wind turbine 17 is installed. The power generation measurement device 17b measures the power generation of the wind turbine 17. The wind turbine operation log recording device 17c records a log of the operating status of the wind turbine 17. The weather observation device 17a, power generation measurement device 17b, and wind turbine operation log recording device 17c of each wind turbine 17 transmit the acquired data to the receiving unit 3.

[0018] The storage unit 5 stores pre-data, data received by the receiving unit 3, and data generated within the wind power generation forecasting device 1. The storage unit 5 includes a weather observation data storage area 5a, a weather forecast data storage area 5b, a power generation measurement data storage area 5c, a wind speed prediction model storage area 5d, a power generation prediction data storage area 5e, a power curve storage area 5f, and a wind turbine operation log storage area 5g. The weather observation data storage area 5a stores weather observation data acquired by the weather observation device 17a of each wind turbine 17 and transmitted to the receiving unit 3. The weather forecast data storage area 5b stores weather forecast data acquired or generated by the weather forecasting device 13 and transmitted to the receiving unit 3. The power generation measurement data storage area 5c stores power generation data for each wind turbine 17 acquired by the power generation measurement device 17b of each wind turbine 17 and transmitted to the receiving unit 3. The wind speed prediction model storage area 5d stores the parameters of the wind speed prediction model generated by the wind speed prediction model learning unit 9, which will be described later. The power generation forecast data storage area 5e stores wind speed forecast data predicted by the wind speed forecast model and power generation forecast data for each wind turbine 17 predicted based on the power curve of each wind turbine 17. The power curve storage area 5f stores the power curve of each wind turbine 17. Here, the power curve is the specification value of the power generation output for each wind turbine 17 in relation to the wind speed, and is generally represented by a graph of wind power versus power generation. The wind turbine operation log storage area 5g stores the log of the operating status of each wind turbine 17, which is acquired by the wind turbine operation log recording device 17c of each wind turbine 17 and transmitted to the receiving unit 3.

[0019] The data processing unit 7 processes the data in the memory unit 5 for use in training and inference of the wind speed prediction model. Specifically, it processes data such as the weather observation data storage area 5a and the weather forecast data storage area 5b, including data format conversion and interpolation, so that the data can be used as input data and training data suitable for AI training and inference.

[0020] The wind speed prediction model learning unit 9 uses the data from the storage unit 5, which has been processed by the data processing unit 7, to machine-learn and generate a wind speed prediction model that predicts the wind speed at the installation locations of multiple wind turbines 17, the details of which will be described later.

[0021] The power generation calculation unit 11 infers and calculates the wind speed at the installation locations of multiple wind turbines 17 using a wind speed prediction model, and calculates the power generation of the multiple wind turbines 17 based on the power curve of each wind turbine 17. The power generation calculation unit 11 includes a wind speed prediction calculation unit 11a and a wind speed-to-power generation conversion unit 11b. The wind speed prediction calculation unit 11a reads a learned wind speed prediction model from the wind speed prediction model storage area 5d of the storage unit 5 and uses it to predict and calculate the future wind speed at the installation locations of the wind turbines 17. The wind speed-to-power generation conversion unit 11b reads the power curve from the power curve storage area 5f of the storage unit 5 and the wind turbine operation log from the wind turbine operation log storage area 5g, and calculates the power generation using the wind speed prediction value calculated by the wind speed prediction calculation unit 11a.

[0022] The output unit 15 outputs power generation forecast data and other data contained in the storage unit 5 for use by other devices. This use includes power generation forecasts for bidding on the spot market, as described later, and use in an EMS (Energy Management System).

[0023] The wind power generation forecasting device 1 is an information processing device such as a personal computer or a server configured on the cloud. Among the components of the wind power generation forecasting device 1, the receiving unit 3, data processing unit 7, wind speed forecasting model learning unit 9, power generation calculation unit 11, and output unit 15 may be software or programs executed by processors such as a CPU, GPU (Graphics Processing Unit), and TPU (Tensor Processing Unit) within the information processing device. The storage unit 5 may be implemented by storage devices such as semiconductor memory or magnetic disks provided inside or outside the information processing device.

[0024] The operation of the wind power generation forecasting device 1 configured in this way will be explained in detail. Here, the operation and function will be explained with the understanding that the power generation forecast will be used to conduct electricity buying and selling bids in the spot market, which is a wholesale market for generated electricity. The spot market is a market in which the buying and selling volume of electricity [kWh] is traded for each of the 48 time slots, which are 30-minute intervals, obtained at 10:00 AM (Japan Standard Time) for the 24 hours of the following day. If the actual amount of electricity generated / consumed in the market is in surplus / deficit, it is necessary to settle the imbalance charge according to the amount of surplus / deficit electricity, and it is important to forecast the amount of power generation appropriately so that the surplus / deficit is as small as possible.

[0025] The weather forecasting device 13 receives weather forecast data or creates it using its own weather forecasting model. The weather forecast data can primarily utilize the Meso-Scale Model (GPV) and Meso-Scale Ensemble Prediction System (MEPS) meso-scale numerical weather prediction models issued by the Japan Meteorological Agency. The mesh points used for data acquisition should be selected to include the point closest to the target wind power plant and its surroundings. MEPS provides forecast results for 21 members using an ensemble method, allowing for a probabilistic capture of weather phenomena compared to the deterministic MSM. However, as will be discussed later, differences exist in the initial time, the number of daily deliveries, and the maximum forecast time, requiring the data to complement each other to construct optimal input data.

[0026] At the wind turbine 17, while generating wind power, the weather observation device 17a observes, measures, and records weather information data at the location where the wind turbine is installed, the power generation measurement device 17b records actual power generation data for each wind turbine 17, and the wind turbine operation log recording device 17c records log data of the operating status of the wind turbine 17. The weather information observed by the weather observation device 17a includes wind speed information, and may also include wind direction, temperature, atmospheric pressure, etc.

[0027] All of the aforementioned data are sent to the receiving unit 3 of the wind power generation forecasting device 1 and stored in the corresponding area of ​​the storage unit 5. In the storage unit 5, the power curve, which is the specification value of the wind speed-power generation output of the wind turbine, is also stored in the power curve storage area 5f as prior information.

[0028] The data processing unit 7 processes the data in the storage unit 5 for use in training and inference of the wind speed prediction model. To ultimately derive prediction data every 30 minutes, it creates representative values ​​every 30 minutes from meteorological observation data stored at some sampling interval. The data that needs processing and the content of the processing are determined by the training and inference method of the wind speed prediction model. Here, representative values ​​may be statistical values ​​of ensemble meteorological data or values ​​of some or all members of the ensemble meteorological data. Here, the example of creating representative values ​​every 30 minutes is explained, but it is not limited to this; for example, representative values ​​may be created every 3 hours, or the input data may be used as is.

[0029] The wind speed prediction model learning unit 9 trains the wind speed prediction model. Details will be described later. In the power generation calculation unit 11, first, the wind speed prediction calculation unit 11a obtains wind speed prediction values ​​for each location where the wind turbines are installed and for each of the 48 time slots covered by the spot market, using a trained wind speed prediction model. Next, the wind speed-power generation conversion unit 11b calculates the power generation amount using the power curve, wind turbine operation log, wind speed prediction values, etc. By fitting the wind speed prediction values ​​to the power curve, the power generation capacity [kW] is calculated for each wind turbine 17, and in the case of a 30-minute time slot, the power generation amount [kWh] can be calculated by multiplying this power generation capacity by 0.5 [h]. At this time, if it is expected that the wind turbines will be shut down in the future by referring to the wind turbine operation log, the power generation amount is treated as 0.

[0030] Finally, the output unit 15 outputs power generation forecast data and other data contained in the storage unit 5 for use in other devices. For example, creating trading plans in the spot market is one example of using the output data.

[0031] <Training a wind speed prediction model> Next, we will explain in detail the training process for the wind speed prediction model used in the wind power generation prediction device 1 described above. Neural networks are considered as wind speed prediction models, and TSMixer and its extension, TSMixer-Ext, which demonstrate particularly high performance for time series processing, may be used. These prediction models learn to predict the wind speed for the current day and the following day based on the previous day's wind speed information, using the prediction day as a reference. By using TSMixer, high prediction performance can be achieved using only an MLP (MultiLayer Perceptron) without using special neural network structures such as Transformers, and the computational load can also be reduced (Non-Patent Literature 1).

[0032] In the use of ensemble weather data, it is common practice to create an equal number of new forecast values ​​in parallel from a given set of weather forecasts, verify their distribution, and then create a single forecast value by integrating multiple forecast values ​​using some method. However, in this invention, representative values ​​of the ensemble weather data are used to improve the accuracy of a single forecast. Here, representative values ​​may be statistical values ​​of the ensemble weather data or values ​​of some or all members of the ensemble weather data.

[0033] Input and output data can be set as follows, for example. • Input (1): 24-hour wind speed data from the previous day (weather information data, for wind turbine locations) • Input (2): MSM wind speed forecast data for the previous day's 24 hours (weather forecast data, per selected grid point) • Input (3): Representative MEPS wind speed forecast values ​​for the previous day's 24 hours (weather forecast data, per selected grid point) • Input (4): MSM wind speed forecast data for the current day and the following day (48 hours) (weather forecast data, per selected grid point) • Input (5): Representative MEPS wind speed forecast values ​​for the current and next day's 48-hour period (weather forecast data, per selected grid point) Output (1): Wind speed data for the current day and the following 48 hours (at the wind turbine location) Figure 2 is a schematic diagram showing an example configuration of TSMixer-Ext with these elements applied. Here, Feature Mixing (code f1, f2) is performed on MEPS inputs (3) and (5), but this is not limited to this. For example, when inputting statistical quantities such as maximum, minimum, mean, and standard deviation as representative values, if Feature Mixing (code f1, f2) is omitted when it is possible to express features more effectively without it, Feature Mixing (code f1, f2) may be removed. Whether or not to use Feature Mixing (code f1, f2) may be decided as appropriate depending on the input features. Furthermore, for MEPS input (5), if a suitable effect can be obtained, the features of MEPS input (5) may be directly input to the Mixing Layer, as shown by the dashed line D.

[0034] Input (1) is, for example, inputting the wind speed data for N wind turbines over the previous 24 hours, averaging the wind speed values ​​at 10-minute intervals (10 minutes, 20 minutes, and 30 minutes), and inputting it as the 30-minute value. Input (2) is inputting the MSM weather forecast data for the previous 24 hours at the nearest M point where each wind turbine is installed. Input (3) is inputting, for example, the representative value of the MEPS weather forecast data for the previous 24 hours at the nearest point to the wind power plant. As mentioned above, the representative value here may be a statistical value of the ensemble weather data (e.g., maximum value, minimum value, median, mean, variance, standard deviation, etc.) or a value of some or all members of the ensemble weather data.

[0035] Input (4) is input of MSM weather forecast data for the current and following 48 hours at the nearest M location where each wind turbine is installed. Input (5) is input of representative values ​​of MEPS weather forecast data for the current and following 48 hours at the location closest to the wind power plant. Output (1) is output of wind speed data for the current and following 48 hours at the N location where the wind turbine is installed. However, in this embodiment, the wind speed value of the MSM is used by combining the east-west and north-south wind speeds into a single wind speed value (Equation 1). Here, V EW V is the wind speed in the east-west direction. SN This represents the wind speed in the north-south direction. The same synthesis is performed for the wind speed in MEPS.

[0036]

number

[0037] For example, if electricity trading for the next 24 hours is conducted by 10:00 AM on the current day, to forecast wind power for the next 24 hours, it is possible to cover 24 hours of forecast data for the next day by combining the forecast values ​​for 18-39 hours ahead from the MSM forecast with an initial time of 6:00 AM on the current day and the forecast values ​​up to 51 hours ahead from the forecast with an initial time of 9:00 PM the previous day. If only MEPS is used, considering the 4-hour distribution lag, the latest MEPS available at 10:00 AM on the current day is a forecast with an initial time of 3:00 AM on the current day, and the forecast values ​​up to 39 hours ahead from that initial time only provide forecast values ​​up to 6:00 PM the next day. MEPS forecast values ​​can be used as input values ​​as they are. Alternatively, by combining them with MSM as appropriate, all time to be forecasted can be covered as input values, improving the accuracy of learning and inference. Here, MSM delivers forecasts in one-hour intervals of 0 to 39 hours or 78 hours at eight initial times a day: 9:00, 12:00, 15:00, 18:00, 21:00, 24:00, 27:00 (3:00), and 30:00 (6:00). MEPS delivers forecasts in three-hour intervals of 0 to 39 hours at four initial times a day: 9:00, 15:00, 21:00, and 27:00 (3:00). When interpolating data in 30-minute intervals in the data processing unit 7, missing values ​​are filled in with either the previous or subsequent value.

[0038] Regarding the output data, historical weather data will be used as training data during the learning process. In particular, the learning model will be characterized by obtaining future wind speed forecast data for each wind turbine location by inputting historical wind speed data and wind speed forecast data, and by using representative MEPS data (statistical values ​​such as the minimum and maximum values ​​for each time point in the ensemble weather data, or the values ​​of some or all members of the ensemble weather data) listed in inputs (3) and (5).

[0039] The learning procedure applies a general procedure to the learning model used. Once the learning is complete, the model is stored in the wind speed prediction model memory area and referenced by the wind speed prediction calculation unit when calculating power generation.

[0040] As described above, in this embodiment, a wind speed prediction model is generated by machine learning by inputting historical wind speed data and wind speed prediction data. This wind speed prediction model is used to obtain future wind speed prediction data at the locations where each wind turbine is installed, and the amount of power generated is predicted based on the power curve of each wind turbine. For training and inference of the wind speed prediction model, representative values ​​of the ensemble weather forecast data, i.e., statistical values ​​of the ensemble weather data or values ​​of some or all members of the ensemble weather data, are used. Here, the wind speed prediction model uses a machine learning model specialized for time series processing. Therefore, it is possible to obtain highly accurate wind power generation prediction values. Thus, by accurately predicting wind speed, it is possible to provide a wind power generation prediction device and a wind power generation prediction method that can improve the accuracy of wind power generation prediction.

[0041] (modified version) Next, a modified example of this embodiment will be described. In this modified version, in addition to the mean squared error function between the training value and the predicted value, a function that evaluates the reliability of the weather forecast data is used as the learning loss function in the wind speed prediction model learning unit 9. Since the other components are the same, the block diagram in Figure 1 and the configuration of the wind speed prediction model in Figure 2 are the same in this modified version as well.

[0042] This section explains the loss function in this modified example. The loss function is a function used to evaluate the magnitude of the difference (loss) between the target value (actual value, correct answer) and the predicted value output by the model. By minimizing / maximizing this loss value, the machine learning model is optimized. As the first element of the loss function in this modified example, we can consider the mean squared error function between the target value (actual value) and the predicted value. The prediction error, that is, the function used to evaluate the predicted value so that it becomes the most likely predicted value, assuming a Gaussian distribution model between the target value and the predicted value, is expressed as shown in the following equation [Equation 2]. This is an expression of the likelihood of a Gaussian distribution.

[0043]

number

[0044] Here, τ is the index of the predicted time, n is the index of the wind turbine, and y true,n,τ , where y is the measured value which is the target value for wind turbine n at time τ, pred,n,τ is the predicted value of wind turbine n at time τ generated by the model. T is the number of samples at the prediction time, N is the number of wind turbines, and σ is a setting parameter corresponding to the standard deviation of the prediction error. In this modified example, the function in [Equation 2] is converted to a log-likelihood and used. Expressing [Equation 2] as a log-likelihood gives the following equation [Equation 3]. Hereafter, this element will be called Target Loss.

[0045]

number

[0046] To improve prediction accuracy, the model is adjusted so that the value of [Equation 3] is maximized, that is, it increases in the positive direction. In other words, the prediction model is trained so that the score of [Equation 3] is as large as possible for all predicted values. Here, the first and second terms can be considered as constants independent of the prediction model, so to improve prediction accuracy, the parameters of the prediction model should be adjusted so that the third term approaches zero.

[0047] In addition to the Target Loss in [Equation 3], this modified example uses a function for evaluating the reliability of weather forecast data as the second element of the loss function. Assuming that the degree of matching between the representative weather forecast data MSM of the ensemble weather data MEPS and the predicted value is modeled by a Gaussian distribution, this loss function is expressed as the following [Equation 4] as the likelihood of the Gaussian distribution.

[0048]

Mathematical Expression

[0049] Here, τ' is the index of the prediction time, n is the index of the wind turbine, y MSM,τ’ is the value of the representative weather forecast data MSM of the ensemble weather data MEPS at time τ', y pred,n,τ’ is the predicted value of wind turbine n at time τ' generated by the model. T' is the number of samples of prediction time, N is the number of wind turbines, and σ τ’ is the variation in prediction of the ensemble weather data MEPS for each prediction time, that is, a parameter corresponding to the standard deviation. In [Equation 4], since the number of wind turbines is the same as that of Target Loss in [Equation 3], the same symbols n and N are used for wind turbines. Since the forecasting sampling and time length differ between data types for the prediction time index, different symbols τ' and T' from those used for Target Loss are adopted for prediction time. Similar to Target Loss, expressing [Equation 4] as a log-likelihood gives the following [Equation 5]. Hereinafter, this element is referred to as MSM Loss.

[0050]

Mathematical Expression

[0051] Regarding MSM Loss in [Equation 5], in order to improve prediction accuracy, the model is adjusted such that the value of [Equation 5] is maximized, that is, increases in the direction of positive values. Here, σ in [Equation 5] τ’ , for the same error amount, the standard deviation σ τ’A larger value results in a smaller error gradient (the slope of the tangent line between the MSM forecast value and the model's predicted value in a Gaussian distribution), which reduces the effect of gradient descent-based learning on bringing the predicted value closer to the MSM forecast value, meaning that the effect of bringing it closer to the target value becomes relatively larger.

[0052] σ τ’ There are two ways to set this: (1) by providing a fixed value derived from evaluation beforehand, and (2) by dynamically determining it for each training sample. In case (1), the goal is to achieve an effect by balancing Target Loss and MSM Loss, so MEPS has σ at each prediction time. τ’ The value should not deviate significantly from the value of σ. τ’ The value of MEPS's σ τ’ The value is set near the value of and evaluated to find and set a suitable constant. In case (2), for example, the mean of the standard deviation of MEPS for each prediction time may be used for each training sample.

[0053] In this modified version, the loss function is calculated by adding the function of MSM Loss [Equation 5] to the function of Target Loss [Equation 3]. As mentioned above, Target Loss [Equation 3] aims to minimize the mean squared error between the measured value and the predicted value, while MSM Loss [Equation 5] evaluates the discrepancy between the MSM representative value and the predicted value of the ensemble meteorological data MEPS, and σ τ’ However, the larger the value, the more tolerance it allows for predictions that deviate from the MSM representative value.

[0054] As described above, in this modified version, in addition to the mean squared error function between the training value and the predicted value, a function that evaluates the reliability of the weather forecast data is used as the loss function for learning the wind speed prediction model in the above embodiment. Therefore, in addition to the effects and benefits of the above embodiment, it becomes possible to generate a prediction model with further improved prediction accuracy.

[0055] In addition to the above, it is possible to select or replace the configurations listed in the above embodiments, or to change them to other configurations as appropriate, as long as it does not deviate from the spirit of the present invention. [Explanation of Symbols]

[0056] 1. Wind power generation forecasting device 3. Receiving Unit 5 Storage section 5a Weather observation data storage area 5b Weather forecast data storage area 5c Power generation measurement data storage area 5D Wind Speed ​​Prediction Model Memory 5e Power generation forecast data storage area 5f Power curve memory area 5G wind turbine operation log storage area 7. Data Processing Department 9. Wind Speed ​​Prediction Model Learning Section 11. Power generation calculation unit 11a Wind Speed ​​Prediction Calculation Unit 11b Wind speed / power generation conversion unit 13 Weather forecasting device 15 Output section 17 Windmill 17a Weather observation equipment 17b Power generation measurement device 17c Wind turbine operation log recording device

Claims

1. A data processing unit processes weather observation data from multiple wind turbines of a wind power generation facility and weather forecast data including representative values ​​from ensemble weather forecast data acquired from external sources, for the purpose of training and inference of a wind speed prediction model. A wind speed prediction model learning unit generates a wind speed prediction model that predicts the wind speed at the installation locations of the multiple wind turbines using the data processed by the data processing unit, A power generation calculation unit that calculates the amount of power generated by the multiple wind turbines based on the power curve, by inferring and calculating the wind speed at the installation locations of the multiple wind turbines using the wind speed prediction model, A wind power generation forecasting device equipped with the following features.

2. The wind power generation forecasting device according to claim 1, wherein the representative value of the ensemble weather forecast data is a statistical value of the ensemble weather data or the value of some or all members of the ensemble weather data.

3. The wind power generation forecasting device according to claim 1 or 2, wherein, in addition to the mean squared error function between the training value and the predicted value, a function that evaluates the reliability of the weather forecast data is used as the loss function for training the wind speed forecasting model.

4. To train and infer from wind speed prediction models, weather observation data from multiple wind turbines and weather forecast data including representative values ​​from externally acquired ensemble weather forecast data are processed. By processing the data as described above, a wind speed prediction model is generated that uses the processed data to predict the wind speed at the installation locations of the multiple wind turbines. The wind speed at the installation locations of the multiple wind turbines is inferred and calculated using the wind speed prediction model, and the amount of power generated by the multiple wind turbines is calculated based on the power curve. A method for predicting wind power generation, including the following.

5. The wind power generation forecasting method according to claim 4, wherein the representative value of the ensemble weather forecast data is a statistical value of the ensemble weather data or a value of some or all members of the ensemble weather data.

6. The wind power generation prediction method according to any one of claims 4 or 5, wherein, in addition to the mean squared error function between the training value and the predicted value, a function that evaluates the reliability of the weather forecast data is used as the loss function for training the wind speed prediction model.

Images