Short-term wind power combined prediction method based on numerical weather forecast wind speed correction

Abstract

The invention discloses a short-term wind power combined prediction method based on numerical weather forecast wind speed correction, and belongs to the field of wind energy and electric power systems in electrical engineering, and the prediction method comprises the following specific steps: I, collecting and preprocessing actually measured wind speed data and NWP forecast data of a wind power plant, and dividing the data into a training set and a test set; iI, analyzing an error between the NWP wind speed and the actually measured wind speed, including an amplitude and a phase error, and constructing a ResNet-GRU model to correct the NWP wind speed to obtain a corrected wind speed; iII, constructing a wind power prediction model, extracting spatial correlation of meteorological features through the model, capturing a long-term dependency relationship of the features, and highlighting key meteorological factors; according to the method, the accuracy of wind power prediction is further improved, a good data basis is provided for subsequent prediction, the prediction error of the model is remarkably reduced, and a new thought is provided for research on a short-term wind power prediction precision improvement technology.

Description

Technical Field

[0001] The present invention relates to the field of wind energy and electric power systems in electrical engineering, and in particular to a short-term wind power combination prediction method based on numerical weather forecast wind speed correction. Background Art

[0002] Under the background of "dual carbon", accelerating the transformation of energy structure to a low-carbon, clean and renewable energy system and building a new power system with new energy as the main body are important measures to achieve the dual carbon goals. At present, my country's wind power industry has entered a stage of rapid development, the wind power market continues to prosper, and the proportion of wind power connected to the grid continues to grow. However, at this stage, my country's power system dispatching and operation mechanism is not sound, the peak-shaving capacity is insufficient, and it is difficult to adapt to the requirements of large-scale grid connection and absorption of wind power. The problem of wind abandonment is serious in some areas. To effectively solve the problem of wind abandonment and improve the dispatching and operation capacity of the power system, it is inevitable to rely on accurate wind power output forecasting. The accuracy of wind power forecasting directly affects the dispatching optimization results of the power grid.

[0003] At present, according to the modeling mechanism, the technical routes of wind power prediction can be roughly divided into physical methods and statistical learning methods. The physical method establishes fluid mechanics and thermodynamics models based on the topography of the wind farm to solve information such as wind speed and wind direction, and then calculates the wind power output from the wind power curve. Limited by the temporal and spatial resolution, the physical method is generally more suitable for medium- and long-term predictions. The statistical learning method is to perform feature analysis on the historical data of the wind farm to establish a nonlinear mapping relationship between wind power characteristics and prediction results. Due to the rapid development of artificial intelligence in recent years, and with the increase in the amount of collected data and the increase in complexity, many scholars have introduced deep learning algorithms to solve the above problems. Deep learning methods represented by long short-term memory (LSTM), back propagation, Dropout, attention machine (AM), etc. are widely used in prediction tasks.

[0004] In the short-term wind power forecast, the wind power forecast based on numerical weather prediction (NWP) is more in line with the actual forecast scenario and has stronger practical applicability. However, the quality of NWP data has a great impact on the forecast results. It is found in actual operation that there is a certain inherent error between the forecast data of NWP and the measured data. Therefore, in order to reduce this inherent error, many scholars have corrected the NWP wind speed. The main methods used in the prior art are: using variational mode decomposition to decompose the NWP wind speed, and then using a bidirectional gated recurrent unit (GRU) to correct the forecast wind speed; considering the spatial correlation of wind speed and using Gaussian process regression (GPR) to correct the NWP wind speed, improve the correlation between the forecast wind speed and the actual wind speed; analyzing multi-location NWP data, using a time series convolutional neural network to establish a wind speed correction model for multi-location NWP, and improving the wind speed correction effect. However, most of the above correction methods use a single neural network for wind speed correction, the relationship between NWP data and measured data is not fully mined, and the model is prone to problems such as gradient explosion during training.

[0005] For any prediction model, prediction errors are common. And because the prediction ability of a single model is limited, it often leads to problems such as low robustness and weak applicability. Therefore, the combined prediction model gradually shows its advantages. However, although the combined model combines the advantages of a single model, it is also easy to cause the superposition of model errors. And the combined model has increased uncertainty due to its complex network structure, and it is difficult to select the parameters of the prediction model. Therefore, many scholars have improved it by combining a series of optimization algorithms. The main methods used in the prior art are: using the isolation forest algorithm to detect abnormal data, and using the improved Sky Eagle optimization algorithm to optimize the LSTM model parameters, so as to establish the IAO-LSTM model for wind power prediction; using the IGWO algorithm to determine the number of nodes in the hidden layer of the model network and the weight learning rate, and proposing a LightGBM-GRU point prediction model, which has achieved better prediction results than other algorithms. However, the above-mentioned optimization algorithm has a complex structure, slow convergence speed, and is prone to fall into the local optimal solution. Therefore, a more suitable intelligent algorithm needs to be selected to be suitable for optimizing more hyperparameters.

[0006] Wind power time series is mainly affected by wind speed and is highly random and volatile. The data quality of Numerical Weather Prediction (NWP) has a significant impact on the accuracy of short-term wind power forecasts. However, NWP forecast errors are common, posing a challenge to the accuracy of wind power forecasts. Summary of the invention

[0007] The purpose of the present invention is to solve the defects in the prior art and to propose a short-term wind power combination prediction method based on numerical weather forecast wind speed correction.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A short-term wind power combination forecasting method based on wind speed correction of numerical weather forecast is proposed. The specific steps of the forecasting method are as follows:

[0010] Ⅰ. Collect and preprocess the measured wind speed data and NWP forecast data of wind farms, and divide the data into training set and test set;

[0011] Ⅱ. Analyze the error between NWP wind speed and measured wind speed, including amplitude and phase error, and construct a ResNet-GRU model to correct the NWP wind speed to obtain the corrected wind speed;

[0012] III. Construct a wind power prediction model to extract the spatial correlation of meteorological characteristics, capture the long-term dependence of characteristics, and highlight key meteorological factors;

[0013] IV. Use KOA optimization algorithm to adjust the hyperparameters in the wind power prediction model;

[0014] V. Train the adjusted wind power prediction model through the Keras framework, and use the trained model to predict wind power.

[0015] As a further solution of the present invention, the specific steps of constructing the ResNet-GRU model in step II are as follows:

[0016] S1.1: The ResNet structure is used as a feature extractor. A GRU model is created based on the architecture of the input layer, GRU layer, and output layer. The deep features of NWP wind speed are then captured through multiple residual blocks in the ResNet structure.

[0017] S2.1: The features extracted by the ResNet structure are fused with the GRU input features, and the fused features are then input into the GRU model for training. The known real wind speed data and the output of the numerical model are used as supervision signals. The error between the predicted value and the true value is calculated, and the model parameters are optimized according to minimizing the error between the predicted value and the true value. The NWP wind speed is corrected in real time based on the optimized ResNet-GRU model.

[0018] As a further solution of the present invention, the specific steps of constructing the wind power prediction model in step III are as follows:

[0019] S2.1: Based on the CNN, LSTM and Attention mechanism, the CNN-LSTM-Attention model architecture is used to create a corresponding wind power prediction model. The CNN convolution layer is used to extract the local features of the wind speed data. The spatial features of the wind speed data are captured through the designed CNN convolution layer. Then, LSTM modeling is performed and the features extracted by CNN are input into the LSTM network to capture the temporal information of the wind speed data.

[0020] S2.2: After the LSTM layer, the Attention mechanism is used to weight the features of important time steps, and the outputs of the CNN, LSTM, and Attention layers are fused to construct the final prediction layer;

[0021] S2.3: Set the loss function and optimizer, train the model to minimize the prediction error, and use the KOA optimization algorithm during the training process to automatically optimize the learning rate, convolution kernel size, and LSTM unit number hyperparameters in the wind power prediction model.

[0022] As a further solution of the present invention, the specific steps of extracting the spatial correlation of meteorological features, capturing the long-term dependency of features, and highlighting key meteorological factors in step III are as follows:

[0023] S3.1: Divide the data into d days, n data per day, and m meteorological factors per data, forming an n×m×d matrix as the input structure of the CNN model. The CNN convolutional layer generates the output expression based on the input data. The corresponding output:

[0024] S3.2: The CNN pooling layer uses a 2×2 filter and a sliding window for sampling to reduce the data feature size and network parameters. The data is then input to the LSTM layer through the fully connected layer. The input vector is passed through the hidden layer of the LSTM to calculate the intermediate state of the meteorological data. The attention mechanism uses the function Calculate the similarity between the feature vectors of the intermediate state and the hidden state;

[0025] S3.3: The attention weight of the hidden layer vector of the meteorological data is obtained through the softmax function, and the weighted sum is performed with the similarity between the feature vectors of the intermediate state and the hidden state to obtain the output of the attention layer. Finally, the sum result is input into the fully connected layer to obtain the predicted value of wind power. The specific calculation formula of the weighted summation is as follows:

[0026]

[0027]

[0028] As a further solution of the present invention, the specific steps of using the KOA optimization algorithm to adjust the hyperparameters in the wind power prediction model in step IV are as follows:

[0029] S4.1: Determine the learning rate, convolution kernel size, and LSTM unit number hyperparameters and their value ranges that need to be optimized for the wind power prediction model, and construct the KOA mathematical model. The specific expression of the KOA mathematical model is as follows:

[0030]

[0031] S4.2: Select the optimization library based on the constructed KOA mathematical model and then use the minimum MAPE as the objective function , and configure the optimizer, then start the optimization, run the optimization program, optimize the selection of each set of hyperparameters, and continuously adjust the hyperparameters.

[0032] As a further solution of the present invention, the specific steps of training the adjusted wind power prediction model through the Keras framework in step V and using the trained model to predict wind power are as follows:

[0033] S5.1: Initialize the wind power prediction model weight learning rate, convolution kernel size and number of neurons through the KOA algorithm, and use the sigmoid function as the activation function of the model;

[0034] S5.2: The data within the past year is used as the training set data, and the data within 7 days after the training set is cut off is used as the test set data. The wind power prediction model is then trained and predicted by the training set and the test set, and the final prediction result is output.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] This short-term wind power combined prediction method based on numerical weather forecast wind speed correction learns the volatility characteristics of wind speed through the GRU model, and fuses the features extracted by the feature extractor of the ResNet structure with the GRU input features, and inputs the fused features into the GRU for training to optimize the overall model parameters. Then, the KOA intelligent algorithm is integrated to construct a wind power prediction model based on CNN-LSTM-Attention. The CNN in the model extracts the local spatial correlation of meteorological features, and the LSTM captures the long-term dependency of time series features such as wind speed. The Attention mechanism highlights the influence of key factors by assigning weights to different features. Then, the model is trained using the corrected wind speed and historical wind power data, and the hyperparameters of the wind power prediction model are automatically optimized through the KOA intelligent algorithm. The final prediction result is output through the trained wind power prediction model, which can further improve the accuracy of wind power prediction, provide a good data basis for subsequent predictions, significantly reduce the prediction error of the model, and provide new ideas for the research on short-term wind power prediction accuracy improvement technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0038] Figure 1 It is a flowchart of the short-term wind power combination prediction method based on numerical weather forecast wind speed correction proposed by the present invention;

[0039] Figure 2 It is a logical framework diagram of the short-term wind power combination prediction method based on numerical weather forecast wind speed correction proposed by the present invention;

[0040] Figure 3 A comparison diagram of the NWP wind speed and measured wind speed distribution and error curve of the short-term wind power combined prediction method based on numerical weather forecast wind speed correction proposed by the present invention;

[0041] Figure 4 A wind speed correction principle diagram based on the ResNet-GRU model for the short-term wind power combined prediction method based on numerical weather forecast wind speed correction proposed by the present invention;

[0042] Figure 5 This is a structural diagram of the CNN-LSTM-Attention combined model of the short-term wind power combined prediction method based on numerical weather forecast wind speed correction proposed in the present invention. DETAILED DESCRIPTION

[0043] Reference Figure 1-Figure 5, a short-term wind power combination forecasting method based on numerical weather forecast wind speed correction, the specific steps of the forecasting method are as follows:

[0044] The measured wind speed data and NWP forecast data of wind farms are collected and preprocessed, and the data are divided into training sets and test sets.

[0045] It should be further explained that the wind speed and wind power data of the wind farm in the past year are collected. The data type is NWP data: wind speed forecast values ​​at four heights of 10m, 30m, 50m, and 70m, once every 15 minutes, of which the measured data includes the measured wind speed at the hub height of the wind farm and the corresponding wind power output. During the peak load stage, the CO2 required for the high-power operation of the carbon capture power plant can be stored in the solution storage device to reduce the energy consumption of carbon capture. At the same time, the energy storage system releases electricity to meet the load demand. The wind speed predicted by NWP is compared with the measured wind speed, and the amplitude and phase errors of the wind speed are analyzed. In different seasons, the error is manifested as different amplitudes and phase delays, and the ResNet module is used to extract the periodic characteristics of historical data.

[0046] The errors between the NWP wind speed and the measured wind speed are analyzed, including amplitude and phase errors, and a ResNet-GRU model is constructed to correct the NWP wind speed to obtain the corrected wind speed.

[0047] Specifically, refer to Figure 2-Figure 4 It can be seen that the ResNet structure is used as a feature extractor, and a GRU model is created based on the architecture of the input layer, GRU layer, and output layer. Then, the deep features of the NWP wind speed are captured through multiple residual blocks in the ResNet structure. The features extracted by the ResNet structure are fused with the GRU input features, and the fused features are input into the GRU model for training. The known real wind speed data and the output of the numerical model are used as supervision signals, and the root mean square error is used to calculate the wind speed of the NWP. The error between the predicted value and the true value is calculated, and the model parameters are optimized according to minimizing the error between the predicted value and the true value. The NWP wind speed is corrected in real time based on the optimized ResNet-GRU model.

[0048] A wind power prediction model is constructed to extract the spatial correlation of meteorological features, capture the long-term dependencies of features, and highlight key meteorological factors.

[0049] Specifically, based on the CNN, LSTM and Attention mechanism of the CNN-LSTM-Attention model architecture, a corresponding wind power prediction model is created. The CNN convolution layer is used to extract the local features of the wind speed data. The spatial features of the wind speed data are captured by the designed CNN convolution layer. Then LSTM modeling is performed. The features extracted by CNN are input into the LSTM network to capture the timing information of the wind speed data. After the LSTM layer, the Attention mechanism is used to weight the features of important time steps. The outputs of the CNN, LSTM and Attention layers are fused to construct the final prediction layer. The loss function and optimizer are set, and the model is trained to minimize the prediction error. During the training process, the KOA optimization algorithm is used to automatically optimize the learning rate, convolution kernel size and LSTM unit number hyperparameters in the wind power prediction model.

[0050] Specifically, the data is divided into d days, with n data per day and m meteorological factors per data, forming an n×m×d matrix as the input structure of the CNN model. The CNN convolutional layer generates the output expression based on the input data. Corresponding to the output, the CNN pooling layer uses a 2×2 filter and a sliding window for sampling to reduce the data feature size and network parameters, and then inputs the data to the LSTM layer through the fully connected layer. The input vector calculates the intermediate state of the meteorological data through the hidden layer of the LSTM. The attention mechanism uses the function The similarity between the feature vectors of the intermediate state and the hidden state is calculated, and the attention weight of the hidden layer vector of the meteorological data is obtained through the softmax function. The weighted sum is taken with the similarity between the feature vectors of the intermediate state and the hidden state to obtain the output of the attention layer. Finally, the sum result is input into the fully connected layer to obtain the predicted value of wind power.

[0051] It should be further explained that the specific calculation formula of weighted summation is as follows:

[0052]

[0053]

[0054] The KOA optimization algorithm is used to adjust the hyperparameters in the wind power prediction model.

[0055] Specifically, determine the learning rate, convolution kernel size, and LSTM unit number hyperparameters and their value ranges that need to be optimized for the wind power prediction model, build a KOA mathematical model, select the optimization library based on the constructed KOA mathematical model, and then use the minimum MAPE as the objective function. And configure the optimizer, then start the optimization, run the optimization program, optimize and select each set of hyperparameters, and continuously adjust the hyperparameters.

[0056] In addition, it should be noted that the specific expression of the KOA mathematical model is as follows:

[0057]

[0058] The adjusted wind power prediction model is trained through the Keras framework, and the trained model is used for wind power prediction.

[0059] Specifically, refer to Figure 2 as well as Figure 5 It can be seen that the KOA algorithm is used to initialize the weight learning rate, convolution kernel size, and number of neurons of the wind power prediction model, and the sigmoid function is used as the activation function of the model. The data within the past year is used as the training set data, and the data within 7 days after the training set is cut off is used as the test set data. Then, the wind power prediction model is trained and predicted through the training set and the test set, and the final prediction result is output.

Claims

1. A short-term wind power combination forecasting method based on numerical weather forecast wind speed correction, characterized in that: The specific steps of the prediction method are as follows: Ⅰ. Collect and preprocess the measured wind speed data and NWP forecast data of wind farms, and divide the data into training set and test set; Ⅱ. Analyze the error between NWP wind speed and measured wind speed, including amplitude and phase error, and construct a ResNet-GRU model to correct the NWP wind speed to obtain the corrected wind speed; III. Construct a wind power prediction model to extract the spatial correlation of meteorological characteristics, capture the long-term dependence of characteristics, and highlight key meteorological factors; IV. Use KOA optimization algorithm to adjust the hyperparameters in the wind power prediction model; V. Train the adjusted wind power prediction model through the Keras framework, and use the trained model to predict wind power.

2. The short-term wind power combination forecasting method based on numerical weather forecast wind speed correction according to claim 1 is characterized in that: The specific steps for building the ResNet-GRU model described in step II are as follows: S1.1: The ResNet structure is used as a feature extractor. A GRU model is created based on the architecture of the input layer, GRU layer, and output layer. The deep features of NWP wind speed are then captured through multiple residual blocks in the ResNet structure. S2.1: The features extracted by the ResNet structure are fused with the GRU input features, and the fused features are then input into the GRU model for training. The known real wind speed data and the output of the numerical model are used as supervision signals. The error between the predicted value and the true value is calculated, and the model parameters are optimized according to minimizing the error between the predicted value and the true value. The NWP wind speed is corrected in real time based on the optimized ResNet-GRU model.

3. The short-term wind power combination forecasting method based on numerical weather forecast wind speed correction according to claim 1 is characterized in that: The specific steps of constructing the wind power prediction model described in step III are as follows: S2.1: Based on the CNN, LSTM and Attention mechanism, the CNN-LSTM-Attention model architecture is used to create a corresponding wind power prediction model. The CNN convolution layer is used to extract the local features of the wind speed data. The spatial features of the wind speed data are captured through the designed CNN convolution layer. Then, LSTM modeling is performed and the features extracted by CNN are input into the LSTM network to capture the temporal information of the wind speed data. S2.2: After the LSTM layer, the Attention mechanism is used to weight the features of important time steps, and the outputs of the CNN, LSTM, and Attention layers are fused to construct the final prediction layer; S2.3: Set the loss function and optimizer, train the model to minimize the prediction error, and use the KOA optimization algorithm during the training process to automatically optimize the learning rate, convolution kernel size, and LSTM unit number hyperparameters in the wind power prediction model.

4. The short-term wind power combination forecasting method based on numerical weather forecast wind speed correction according to claim 3 is characterized in that: The specific steps of extracting the spatial correlation of meteorological features, capturing the long-term dependencies of features, and highlighting key meteorological factors described in step III are as follows: S3.1: Divide the data into d days, n data per day, and m meteorological factors per data, forming an n×m×d matrix as the input structure of the CNN model. The CNN convolutional layer generates the output expression based on the input data. The corresponding output: S3.2: The CNN pooling layer uses a 2×2 filter and a sliding window for sampling to reduce the data feature size and network parameters. The data is then input to the LSTM layer through the fully connected layer. The input vector is passed through the hidden layer of the LSTM to calculate the intermediate state of the meteorological data. The attention mechanism uses the function Calculate the similarity between the feature vectors of the intermediate state and the hidden state; S3.3: The attention weight of the hidden layer vector of the meteorological data is obtained through the softmax function, and the weighted sum is performed with the similarity between the feature vectors of the intermediate state and the hidden state to obtain the output of the attention layer. Finally, the sum result is input into the fully connected layer to obtain the predicted value of wind power. The specific calculation formula of the weighted summation is as follows:

5. The short-term wind power combination forecasting method based on numerical weather forecast wind speed correction according to claim 4 is characterized in that: The specific steps of using the KOA optimization algorithm to adjust the hyperparameters in the wind power prediction model described in step IV are as follows: S4.1: Determine the learning rate, convolution kernel size, and LSTM unit number hyperparameters and their value ranges that need to be optimized for the wind power prediction model, and construct the KOA mathematical model. The specific expression of the KOA mathematical model is as follows: S4.2: Select the optimization library based on the constructed KOA mathematical model and then use the minimum MAPE as the objective function And configure the optimizer, then start the optimization, run the optimization program, optimize and select each set of hyperparameters, and continuously adjust the hyperparameters.

6. The short-term wind power combination forecasting method based on numerical weather forecast wind speed correction according to claim 5 is characterized in that: The specific steps of training the adjusted wind power prediction model through the Keras framework described in step V and using the trained model to predict wind power are as follows: S5.1: Initialize the wind power prediction model weight learning rate, convolution kernel size and number of neurons through the KOA algorithm, and use the sigmoid function as the activation function of the model; S5.2: The data within the past year is used as the training set data, and the data within 7 days after the training set is cut off is used as the test set data. The wind power prediction model is then trained and predicted by the training set and the test set, and the final prediction result is output.

Images