Multi-meteorological-source fusion wind power short-term power prediction method based on ResNet neural network

Abstract

The invention discloses a multi-meteorological-source fusion wind power short-term power prediction method based on a ResNet neural network. Comprising the steps of numerical weather data acquisition and accumulation, wind power plant operation data acquisition and cleaning, feature processing and sample construction, and construction of a power prediction model based on a ResNet neural network. According to the method, data of multiple groups of numerical weather forecasts are considered, so that the contingency of deviation of a single numerical weather forecast is avoided, and the short-term power prediction precision of the wind power plant can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power, and in particular to a short-term wind power prediction method based on multi-meteorological source fusion using a ResNet neural network. Background Art

[0002] As one of the main renewable energy sources, the confidence level of wind power is quite different from other energy forms. This is mainly because wind power depends to a large extent on weather conditions, and changes in wind speed and direction will cause instability in wind power. Due to the intermittent, random, and uncertain characteristics of wind power grid connection, its impact on the power grid during grid connection is relatively large. Therefore, the power system dispatching department needs to make effective planning and scheduling arrangements. Thus, accurate wind power prediction is particularly important.

[0003] The core difficulty in wind power prediction lies in establishing the relationship between meteorological resource elements and actual power. Currently, there are physical methods, statistical methods, etc. Among them, physical methods do not require historical data and can be obtained based on the wind speed at the hub of the wind turbine and the theoretical power curve of the wind turbine. Since physical methods are calculated from a theoretical perspective, there is a certain deviation between the predicted power result and the actual power. After the wind farm starts generating electricity, actual data can be used to correct the model. There are many methods to correct the model, such as traditional machine learning models, deep neural network models, etc. Different models have different advantages and application scopes. Traditional machine learning has high computational efficiency and low data requirements. Summary of the Invention

[0004] The purpose of the present invention is to provide a short-term wind power prediction method based on multi-meteorological source fusion using a ResNet neural network. When there is a large amount of data at the station, the deep learning model will show higher accuracy.

[0005] To this end, the present invention adopts the following technical solutions:

[0006] A short-term wind power prediction method based on multi-meteorological source fusion using a ResNet neural network, characterized by comprising the following steps:

[0007] (1) Acquisition and accumulation of numerical weather data;

[0008] (2) Acquisition and cleaning of wind farm operation data, where the wind farm operation data includes measured power data and measured wind speed data of the wind farm;

[0009] (3) For the acquired meteorological prediction data and measured power data, perform feature processing and training sample construction; in the feature processing stage, select multiple features of multiple groups of numerical weather forecasts for combination, and add the measured power data at the corresponding moment to form a sample;

[0010] (4) Obtain the data of short-term predicted power, calculate the loss using the actual power in the test set, and constrain the training process of the model to obtain the trained multi-meteorological-source fusion wind power short-term prediction model based on the ResNet neural network;

[0011] (5) In the prediction stage, obtain the predicted meteorological data of multi-meteorological sources, input it into the trained model, and obtain the power prediction data.

[0012] Furthermore, in step (1), obtain multiple sets of numerical weather prediction data. Each set of numerical weather prediction data uses two variables, namely the wind speed at 100m and the wind direction at 100m. Only one set of data with the day-ahead prediction is retained for duplicate data, and all data is interpolated to a specified time interval. This specified time interval is equal to the following specified time interval in step (2), preferably 5 minutes.

[0013] Furthermore, in step (1), the data retained daily is the forecast data generated before 8 o'clock of the previous day.

[0014] Furthermore, in step (1), interpolate the obtained numerical weather data to data at a specified interval. Among them, the interpolation of wind direction data needs to be processed through the change and inverse transformation of meridional wind and zonal wind.

[0015] Furthermore, in step (2), identify the abnormal power generation moments according to the measured wind speed and measured power data, and discard the data at abnormal power generation moments. The measured wind speed and measured power data are the measured power data and measured wind speed data at the minute level (such as 5 minutes or 10 minutes or 15 minutes) of the wind farm.

[0016] Furthermore, in step (2), use the four-parameter Logistics curve fitting method, quartile method or isolation forest method to identify the abnormal power generation data of the wind farm, discard the data at abnormal moments such as power curtailment and maintenance moments, and interpolate the normal data into data at a specified interval using the linear interpolation algorithm.

[0017] Furthermore, in step (3), uniformly interpolate the data into data at a specified interval, align them according to time, and divide the training set, validation set, and test set according to a ratio. For example, divide the training set, validation set, and test set according to a ratio of 7:2:1.

[0018] Furthermore, sort the processed sample data in ascending order of time, and intercept the latest 10% as the test set. For the remaining 90% of the samples, extract 70% of the test set data and 20% of the validation set data through random sampling;

[0019] Further, in step (4), the constructed ResNet neural network includes 2-4 BasicBlock blocks; the input dimension of the first BasicBlock is twice the number of meteorological sources, the output dimension of the last BasicBlock block is 1, and the input of the intermediate BasicBlock is the same as the output of the previous BasicBlock. Each BasicBlock is composed of three layers of neural network and a residual connection module, where the dimension of the middle layer is twice the dimension of the input layer, and the activation function after the residual connection uses the Tanh function.

[0020] By considering the data of multiple numerical weather forecasts, the present invention avoids the contingency of the deviation of a single numerical weather forecast and can effectively improve the short-term power prediction accuracy of the wind farm. Description of the Drawings

[0021] Figure 1 is the flow chart of the sample data processing described in the present invention.

[0022] Figure 2 From left to right are the schematic diagrams of identifying power data by using the four-parameter Logistics curve fitting method, the quartile method and the Isolation Forest method respectively.

[0023] Figure 3 is the block diagram of the ResNet neural network structure described in the present invention. Detailed Embodiments

[0024] Referring to the attached Figures 1-3 The multi-meteorological source fusion wind power short-term power prediction method based on the ResNet neural network provided by the present invention includes the following steps:

[0025] Step S1: Acquisition and accumulation of numerical weather data. Select 2-3 data sources of numerical weather forecasts, make a data request before 8 o'clock every day, and retain the forecast data for the second day (16h - 40h) in the future. The data should include the wind speed and wind direction at 100m, continuously accumulate data for more than 6 months, and store it in the MySQL database;

[0026] Step S2: Acquisition and cleaning of wind farm operation data. Obtain the measured power data and measured wind speed data of the wind farm at the minute level (5 minutes or 10 minutes or 15 minutes) for more than 6 months, and identify abnormal power generation moments according to the measured wind speed and measured power data, and discard the data at abnormal power generation moments;

[0027] Step S3: Feature processing and sample construction. For the obtained meteorological prediction data and measured power data, the data are uniformly interpolated into data at 5-minute intervals and aligned according to time. Each record includes 01_100ws, 01_100wd, 02_100ws, 02_100wd, …, actual_power, and the training set, validation set, and test set are divided according to the ratio of 7:2:1.

[0028] Step S4: Input the training data set in S3 into the constructed power prediction model based on the ResNet neural network to obtain the data of short-term predicted power, and calculate the loss using the actual power in the test set to constrain the training process of the model, and obtain the trained multi-meteorological source fusion wind power short-term power prediction model based on the ResNet neural network.

[0029] Further, Step S1 specifically includes: obtaining multiple groups of publicly available numerical weather prediction data, and setting a scheduled task to obtain data for a future period before 8 o'clock every day. Figure 1 The schematic diagram of data acquisition is given, where data is acquired at 8 o'clock every day, and the data of the second day (the dark part in the figure) is intercepted. The acquired data is interpolated, and the data is uniformly interpolated into a 5-minute time interval, and the data from 00:00 to 24:00 of the second day is intercepted and saved to the MySQL database, a total of 288 pieces of data.

[0030] In the data interpolation process of Step S1, in order to ensure the correctness of the wind direction interpolation algorithm, first convert the wind speed W s , wind direction W d into the meridional wind U a and zonal wind V a , and the calculation formulas are as follows:

[0031] U a = -W s × sinW d ,

[0032] V a = -W s × cos W d .

[0033] Adopt the linear interpolation algorithm to interpolate the meridional wind and zonal wind into data at 5-minute intervals, and finally convert the meridional wind and zonal wind data into wind speed and wind direction data, and the calculation formulas are as follows:

[0034]

[0035] W d = π + arctan(U a , V a ).

[0036] Further, step S2 specifically includes: obtaining the measured power data and measured wind speed data of the wind farm at the minute level (5 minutes or 10 minutes or 15 minutes) for more than 6 months from the wind farm SCADA system, and using the four-parameter Logistics curve fitting method, quartile method or isolation forest method to identify the abnormal power generation data of the wind farm, and discarding the data at abnormal times such as power curtailment and maintenance times. The recognition effect is as Figure 2 shown, where the light-colored points in the figure are normal power generation data, and the dark-colored points are abnormal power generation data that need to be discarded.

[0037] In step S2, if the measured power data is not at 5-minute intervals, the linear interpolation algorithm is used to interpolate the data into 5-minute time intervals, with 288 records per day, and save them into the MySQL database.

[0038] Further, step S3 specifically includes: obtaining the numerical weather prediction data and measured power data processed in steps S1 and S2, and combining them one by one according to time. After combination, there will be 288 samples per day, and each record contains 01_100ws, 01_100wd, 02_100ws, 02_100wd,..., actual_power. The following table gives a schematic diagram of the sorted samples.

[0039]

[0040] In step S3, the processed sample data is sorted in ascending order of time, and the latest 10% is intercepted as the test set. The remaining 90% of the samples are randomly sampled to extract 70% of the test data and 20% of the validation data.

[0041] Figure 3 The ResNet neural network structure for wind power prediction is given. Further, in step S4: the power prediction model based on the ResNet neural network includes 2 - 4 BasicBlock blocks. The input dimension of the first BasicBlock is twice the number of meteorological sources, the output dimension of the last BasicBlock block is 1, and the input of the intermediate BasicBlock is the same as the output of the previous BasicBlock. Each BasicBlock is composed of three layers of neural network and a residual connection module, where the dimension of the middle layer is twice the dimension of the input layer. The activation function after the residual connection uses the Tanh function.

[0042] In step S4, the training process uses the Adam optimizer and combines a linear learning decay rate. The loss function uses L1Loss, and the formula is as follows:

[0043]

[0044] Where P a is the actual power in the sample, and P t is the short-term predicted power of the model. After training is completed, the final model is obtained.

Claims

1. A short-term wind power prediction method based on the ResNet neural network with multi-meteorological source fusion, characterized in that: It includes the following steps: (1) Acquisition and accumulation of numerical weather data; (2) Acquisition and cleaning of wind farm operation data, where the wind farm operation data includes measured power data and measured wind speed data of the wind farm; (3) Feature processing and training sample construction for the acquired meteorological prediction data and measured power data; (4) Obtain the data of short-term predicted power, calculate the loss using the actual power in the test set, constrain the training process of the model, and obtain the trained short-term wind power prediction model based on the ResNet neural network that fuses multiple meteorological sources; (5) In the prediction stage, obtain the predicted meteorological data of multiple meteorological sources, input it into the trained model, and obtain the power prediction data.

2. The prediction method according to claim 1, wherein: In step (1), multiple groups of numerical weather forecast data are obtained. Each group of numerical weather forecast data uses two variables, namely wind speed at 100m and wind direction at 100m. Only one set of data predicted on the day before is retained for duplicate data, and all data is interpolated to the specified time interval.

3. The prediction method according to claim 2, characterized in that: The data retained daily is the forecast data generated before 8 o'clock on the previous day.

4. The prediction method according to claim 2, wherein: Interpolate the acquired numerical weather data to the data with the specified interval. Among them, the interpolation of wind direction data needs to be processed through the change and inverse transformation of meridional wind and zonal wind.

5. The prediction method according to claim 1, wherein: In step (2), identify the abnormal power generation moments according to the measured wind speed and measured power data, and discard the data at abnormal power generation moments.

6. The prediction method according to claim 5, wherein: Use the four-parameter Logistics curve fitting method, quartile method or isolation forest method to identify the abnormal power generation data of the wind farm, discard the data at abnormal moments, and interpolate the normal data into the data with the specified interval using the linear interpolation algorithm.

7. The prediction method according to claim 1, wherein: In step (3), uniformly interpolate the data into the data with the specified interval, align it according to time, and divide it into a training set, a validation set, and a test set according to a ratio.

8. The prediction method according to claim 7, characterized in that: The processed sample data is sorted in ascending order of time, and the latest 10% is intercepted as the test set. The remaining 90% of the samples are randomly sampled to extract 70% of the test set data and 20% of the validation set data.

9. The prediction method according to claim 1, wherein: The constructed ResNet neural network includes 2 - 4 BasicBlock blocks; the input dimension of the first BasicBlock is twice the number of meteorological sources, the output dimension of the last BasicBlock block is 1, and the input of the intermediate BasicBlock is the same as the output of the previous BasicBlock.

10. The prediction method according to claim 9, wherein: Each BasicBlock is composed of three layers of neural network and a residual connection module. Among them, the dimension of the middle layer is twice the dimension of the input layer, and the activation function after the residual connection uses the Tanh function.

Images