A method for improving the accuracy of wind power prediction
By filtering and processing wind power data and training a model using a long short-term memory network, the problems of large computational load, long time consumption and low accuracy of existing wind power prediction methods are solved, and efficient wind power prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SANMENXIA POWER SUPPLY COMPANY OF STATE GRID HENAN ELECTRIC POWER
- Filing Date
- 2022-08-24
- Publication Date
- 2026-06-23
Smart Images
Figure CN115425645B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wind power forecasting technology, and specifically relates to a method for improving the accuracy of wind power forecasting. Background Technology
[0002] With the proposal of "carbon peaking" and "carbon neutrality" goals, the large-scale grid connection of new energy sources such as wind power is an inevitable trend. In order to avoid difficulties in grid connection of new energy sources such as wind power due to insufficient grid dispatching, it is necessary to predict wind power, which has transient, random and uncertain characteristics. Improving the accuracy of prediction can effectively improve the planning of grid dispatching, which is of great significance for improving the grid connection rate of new energy sources such as wind power.
[0003] Currently, commonly used wind power forecasting methods mainly include numerical weather prediction combined with physical methods, statistical methods relying on wind power signals, and methods based on numerical weather prediction combined with artificial intelligence. Methods based on numerical weather prediction combined with physical methods rely heavily on the physical setup process and are highly dependent on data such as the location and size of wind turbines, posing a challenge for regional-level wind power forecasting. Statistical methods relying on wind power signals decompose one-dimensional time-series wind power signals into multi-dimensional samples through feature value decomposition before prediction. However, since the input samples consist of data from several time points, it is difficult to represent the current pattern of the system, resulting in generally low prediction accuracy. Methods based on numerical weather prediction combined with artificial intelligence are a mainstream approach, but they require a large amount of data for training, leading to high demands on computer hardware and time-consuming training processes. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and propose a method to improve the accuracy of wind power prediction. This method can balance the computational load and time consumption of model training with the prediction accuracy, resulting in low computational load, short time consumption, high prediction accuracy, and high application value.
[0005] The objective of this invention is achieved as follows: a method for improving the accuracy of wind power forecasting, comprising the following steps:
[0006] Step 1: Select numerical weather forecast data, wind turbine operation data, and measured power data for a certain time period;
[0007] Step 2: Perform anomaly processing on the numerical weather forecast data, wind turbine operation data, and measured power data from Step 1;
[0008] Step 3: Normalize the data processed in Step 2;
[0009] Step 4: Divide the normalized data in Step 3 into a training dataset and a test dataset. Select numerical weather forecast data and wind turbine operation data in the training dataset as input data, and select measured power data in the training dataset as output data. Train the training dataset and obtain the baseline prediction accuracy η0 through the test dataset.
[0010] Step 5: Subtract one variable x from the numerical weather forecast data and the wind turbine operation data in turn from the training dataset obtained in Step 4. i Then, using the measured power data from the training dataset as input data, the model is trained again on the training dataset. The resulting model is then tested, and the prediction accuracy η is calculated. i ;
[0011] Step 6: Filter out variables that have a significant impact on prediction accuracy;
[0012] Step 7: Train the long-term data based on the selected variables to obtain a wind power prediction model with a shorter training time and higher prediction accuracy.
[0013] Step 8: Use the variables selected in Step 6 within the time period to be predicted as input variables for the trained model to predict the wind power forecast value for the prediction period.
[0014] The numerical weather forecast data in step 1 includes, but is not limited to, wind speed, wind direction, total irradiance, direct normal irradiance, diffuse horizontal irradiance, air temperature, relative humidity, and air pressure at the hub height of the wind turbine at 10m, 30m, 50m, and 70m (when the hub height is not equal to 70m). The wind turbine operating data includes, but is not limited to, blade pitch angle, nacelle direction, temperature inside the turbine nacelle, impeller diameter, and blade pitch.
[0015] Step 2 specifically includes: processing the abnormal data in the numerical weather forecast data, wind turbine operation data, and measured power data in Step 1 using the nearest neighbor mean imputation method or the similar day missing data imputation method to obtain the processed data.
[0016] Step 3 specifically includes: normalizing the data processed in step 2 to obtain normalized data in the interval [-1, 1]. The normalization method is as follows.
[0017]
[0018] Where x imax and x imin To select a variable x within a time period i The maximum and minimum values of .
[0019] Step 4 specifically includes: dividing the normalized data from step 3 into a training dataset and a test dataset, with the training dataset accounting for [50% to 98%] of the normalized data and the remaining data being the test dataset; selecting numerical weather forecast data and wind turbine operation data from the training dataset as input data; selecting measured power data from the training dataset as output data; training the training dataset using a long short-term memory network; testing the obtained model using the test dataset; and calculating the baseline prediction accuracy η0.
[0020] Step 5 specifically includes: successively reducing a variable x from the numerical weather forecast data and the wind turbine operation data in the training dataset obtained in step 4. i Then, using the measured power data from the training dataset as input data, the model is trained again on the training dataset using a Long Short-Term Memory network. The resulting model is then processed by removing and reducing the variable x. i The prediction accuracy η was calculated by testing the subsequent test dataset. i .
[0021] Step 6 specifically includes: by passing each η i Compared with η0, variables that have a significant impact on prediction accuracy are selected, and the following judgment calculation is used.
[0022]
[0023] If we reduce variable x i Δη after i If the value is greater than the threshold Δη, the variable is considered to have a significant impact on the prediction accuracy; otherwise, it is considered to have a minor impact on the prediction accuracy.
[0024] Step 7 specifically includes: based on the variables that have a significant impact on prediction accuracy selected from numerical weather forecast data and wind turbine operation data in step 6, selecting variables with a longer time period and measured power data, completing the abnormal data processing in step 2 and the data normalization processing in step 3 to obtain normalized data; training the normalized data through a long short-term memory network to obtain a trained model.
[0025] The beneficial effects of this invention are as follows: The method for improving wind power forecasting accuracy, as described in this invention, first selects numerical weather forecast data, wind turbine operation data, and measured power data over a short time period. Through anomaly data processing and data normalization, and by progressively reducing input variables, the impact of variable loss on forecasting accuracy is compared to identify variables with a significant impact on forecasting accuracy. Then, variables and measured power data over a longer time period are selected, and anomaly data processing and data normalization are completed. The normalized data is trained using a Long Short-Term Memory (LSTM) network to obtain a trained model, which is then used for wind power forecasting in actual systems. This method for improving wind power forecasting accuracy strikes a balance between computational load, computation time, and forecasting accuracy during model training, resulting in low computational load, short training time, high forecasting accuracy, and high application value. Attached Figure Description
[0026] Figure 1 This is a flowchart illustrating a method for improving the accuracy of wind power forecasting according to the present invention. Detailed Implementation
[0027] The present invention will now be further described with reference to the accompanying drawings.
[0028] like Figure 1 As shown, a method for improving the accuracy of wind power forecasting includes the following steps:
[0029] Step 1: Select numerical weather forecast data, wind turbine operation data, and measured power data for a certain time period. The numerical weather forecast data includes, but is not limited to, wind speed, wind direction, total irradiance, direct normal irradiance, horizontal diffuse irradiance, air temperature, relative humidity, and air pressure at 10m, 30m, 50m, 70m and at the hub height of the wind turbine (when the hub height is not equal to 70m). The wind turbine operation data includes, but is not limited to, blade pitch angle, nacelle direction, temperature inside the turbine nacelle, impeller diameter, and blade pitch.
[0030] In this embodiment, numerical weather forecast data, wind turbine operation data, and measured power data with a length of 100 days and an interval of 15 minutes are selected. The numerical weather forecast data includes wind speed, wind direction and the angle between the turbine nacelle and the surrounding environment temperature, etc. The wind turbine operation data includes the yaw angle of the nacelle, the temperature inside the turbine nacelle, the pitch angle of blade 1, the pitch angle of blade 2 and the pitch angle of blade 3.
[0031] Step 2: Abnormal data in the numerical weather forecast data, wind turbine operation data, and measured power data in Step 1 are processed using the nearest neighbor mean imputation method or the similar day missing data imputation method to obtain the processed data;
[0032] Step 3: Normalize the data processed in Step 2 to obtain normalized data in the interval [-1, 1]. The normalization method is as follows:
[0033]
[0034] Where x imax and x imin To select a variable x within a time period i The maximum and minimum values;
[0035] Step 4: Divide the normalized data from Step 3 into a training dataset and a test dataset. The training dataset comprises approximately [50%, 98%] of the normalized data, and the remaining data serves as the test dataset. Numerical weather forecast data and wind turbine operation data from the training dataset are selected as input data, and measured power from the training dataset is selected as output data. A Long Short-Term Memory (LSTM) network is used to train the training dataset. The resulting model is then tested using the test dataset, and the baseline prediction accuracy η0 is calculated. In this embodiment, the training dataset comprises 90% of the normalized data, and the resulting baseline prediction accuracy η0 = 0.825.
[0036] Step 5: Subtract one variable x from the numerical weather forecast data and the wind turbine operation data in turn from the training dataset obtained in Step 4. i Then, using the measured power from the training dataset as input data, the model is trained again using a Long Short-Term Memory network. The resulting model is then processed by removing the variable x from the previous step. i The prediction accuracy η was calculated by testing the subsequent test dataset. i In this embodiment, the prediction accuracies for wind speed, wind direction and the angle between the turbine nacelle position, ambient temperature, nacelle yaw angle, turbine nacelle temperature, pitch angle of blade 1, pitch angle of blade 2 and pitch angle of blade 3 are 0.627, 0.728, 0.781, 0.796, 0.765, 0.797, 0.790 and 0.787, respectively.
[0037] Step 6: By passing each η i Compared with η0, variables that have a significant impact on prediction accuracy are selected, and the following judgment calculation is used:
[0038]
[0039] If we reduce variable x i Δη after i If the value is greater than the threshold Δη, the variable is considered to have a significant impact on the prediction accuracy; otherwise, the variable is considered to have a relatively small impact on the prediction accuracy.
[0040] In this embodiment, Δη = 4.5% is selected to identify the variables that have a significant impact on prediction accuracy as wind speed, the angle between wind direction and turbine nacelle position, ambient temperature, and nacelle yaw angle.
[0041] Step 7: Based on the variables that have a significant impact on prediction accuracy, selected from numerical weather forecast data and wind turbine operation data in Step 6, namely wind speed, the angle between wind direction and turbine nacelle position, ambient temperature, and nacelle yaw angle, select variables with a longer time period and measured power data, complete the abnormal data processing in Step 2 and the data normalization processing in Step 3 to obtain normalized data; train the normalized data through a long short-term memory network to obtain the trained model;
[0042] Step 8: Use the variables selected in Step 6 within the time period to be predicted as input variables for the trained model to predict the wind power forecast value for the prediction period.
Claims
1. A method for improving the accuracy of wind power forecasting, characterized in that, It includes the following steps: Step 1: Select numerical weather forecast data, wind turbine operation data, and measured power data for a certain time period; Step 2: Perform anomaly processing on the numerical weather forecast data, wind turbine operation data, and measured power data from Step 1; Step 3: Normalize the data processed in Step 2; Step 4: Divide the normalized data from Step 3 into a training dataset and a test dataset. Select numerical weather prediction data and wind turbine operation data from the training dataset as input data, and select measured power data from the training dataset as output data. Train the training dataset and obtain the baseline prediction accuracy using the test dataset. ; Step 5: Subtract one variable from the numerical weather forecast data and the wind turbine operation data in turn from the training dataset obtained in Step 4. Then, using the measured power data from the training dataset as input data, the model is trained again on the training dataset. The resulting model is then tested, and the prediction accuracy is calculated. ; Step 6: Filter out variables that have a significant impact on prediction accuracy; Step 7: Train the long-term data based on the selected variables to obtain a wind power prediction model with a shorter training time and higher prediction accuracy. Step 8: Use the variables selected in Step 6 within the time period to be predicted as input variables for the trained model to predict the wind power forecast value for the prediction time period. Step 2 specifically includes: processing the abnormal data in the numerical weather forecast data, wind turbine operation data and measured power data in step 1 using the nearest neighbor mean imputation method or the similar day missing data imputation method to obtain the processed data; Step 4 specifically includes: dividing the normalized data from step 3 into a training dataset and a test dataset, with the training dataset comprising [50%, 98%] of the normalized data and the remaining data serving as the test dataset; selecting numerical weather prediction data and wind turbine operation data from the training dataset as input data; selecting measured power data from the training dataset as output data; training the training dataset using a Long Short-Term Memory (LSTM) network; testing the obtained model using the test dataset; and calculating the baseline prediction accuracy. .
2. The method for improving wind power forecasting accuracy as described in claim 1, characterized in that, The numerical weather forecast data in step 1 includes, but is not limited to, wind speed, wind direction, total irradiance, direct normal irradiance, diffuse horizontal irradiance, air temperature, relative humidity, and air pressure at the hub height of the wind turbine at 10m, 30m, 50m, and 70m (when the hub height is not equal to 70m). The wind turbine operating data includes, but is not limited to, blade pitch angle, nacelle direction, temperature inside the turbine nacelle, impeller diameter, and blade pitch.
3. The method for improving wind power forecasting accuracy as described in claim 1, characterized in that, Step 3 specifically includes: normalizing the data processed in step 2 to obtain normalized data in the interval [-1, 1]. The normalization method is as follows. in and To select a variable within a time period The maximum and minimum values of .
4. The method for improving wind power forecasting accuracy as described in claim 1, characterized in that, Step 5 specifically includes: successively reducing one variable from the numerical weather forecast data and the wind turbine operation data in the training dataset obtained in step 4. Then, using the measured power data from the training dataset as input data, the model is trained again on the training dataset using a Long Short-Term Memory network. The resulting model is then processed by removing variables. The prediction accuracy was calculated by testing the subsequent test dataset. .
5. The method for improving wind power forecasting accuracy as described in claim 1, characterized in that, Step 6 specifically includes: passing each and By comparing and selecting variables that have a significant impact on prediction accuracy, the following judgment calculation is used. If reduce variables After Greater than the threshold This means that the variable is considered to have a significant impact on prediction accuracy, or conversely, the variable is considered to have a relatively small impact on prediction accuracy.
6. The method for improving wind power forecasting accuracy as described in claim 1, characterized in that, Step 7 specifically includes: based on the variables that have a significant impact on prediction accuracy selected from numerical weather forecast data and wind turbine operation data in step 6, selecting variables with a longer time period and measured power data, completing the abnormal data processing in step 2 and the data normalization processing in step 3 to obtain normalized data; training the normalized data through a long short-term memory network to obtain a trained model.