A wind power probability prediction method based on data enhancement and CNN-LSTM

Abstract

The application discloses a wind power probability prediction method based on data enhancement and CNN-LSTM, which divides collected NWP data and corresponding power data into two independent data sets for model training and testing; first, the training data set is used to train the Wasserstein GAN to generate a large number of data similar to the training data set samples, and meanwhile, the wind farm data with similar error distribution is used to realize data migration; the real training data set samples and the generated samples are respectively used as the input of the LSTM and CNN models, and a CNN-LSTM combined probability prediction model is formed through a full connection layer; in order to judge the performance of the prediction model, the most suitable prediction model is selected, the test data set is input into the selected CNN-LSTM probability prediction model for prediction, and the output result is obtained. The method is suitable for wind power probability prediction in the scene of newly-built wind farms or insufficient data under extreme weather, has high prediction accuracy compared with direct modeling, has wide data application range, and has popularization value.

Description

Technical Field

[0001] This invention relates to a wind power probabilistic prediction method based on data augmentation and CNN-LSTM, belonging to the field of new energy power prediction. Background Technology

[0002] Building a new power system with new energy as the main body, wind power generation is one of the important forms of new energy power generation, and the development of wind power generation is of great significance to power production. However, due to the large intermittency and volatility of wind power, as well as its randomness, large-scale grid connection of wind power will bring certain problems to the safety and stability of power system operation. Therefore, establishing a complete wind power probability prediction system is of great significance to power system dispatch and safe operation, and can also improve the wind curtailment phenomenon and help the economic operation of the power system. In recent years, with the development of big data artificial intelligence technology and deep learning algorithms, the application of deep learning technology in wind power probability prediction has become possible. Compared with shallow machine learning, the advantages of deep learning are: (1) it can approximate the complex nonlinear function of actual needs through a nonlinear structure trained by multiple network layers; (2) the method of fine-tuning after training the network layer by layer can effectively extract the abstract concept of the problem and effectively solve the problem of local optimal solutions in traditional shallow neural networks; (3) the network structure is flexible and can complete supervised learning under different structures, with better adaptability.

[0003] For wind power probability prediction, there are currently two main prediction methods: parametric and non-parametric. Non-parametric methods can better reflect the probability distribution of wind power because they do not require prior assumptions that wind power follows a specific probability distribution. With the commissioning of a large number of newly built wind farms, how to perform deep learning modeling of wind power under insufficient sample conditions has become a research hotspot. Currently, the main methods for deep learning modeling of newly built wind farms are: (1) using generative models such as autoencoders and generative adversarial networks to expand the sample; (2) using transfer learning methods to transfer the complete wind farm data or its wind power prediction model to the newly built wind farm. When the sample is scarce, using generative adversarial networks can effectively expand the sample based on the existing sample and achieve the effect of data augmentation. However, for time series data such as wind power, generative adversarial networks cannot learn the temporal correlation within the data very well, and further improvements to the prediction method are needed. Summary of the Invention

[0004] The purpose of this invention is to provide a wind power probability prediction method based on data augmentation and CNN-LSTM to solve the problems existing in the above-mentioned background technology.

[0005] The objective of this invention is achieved by the following technical measures:

[0006] A wind power probability prediction method based on data augmentation and CNN-LSTM is characterized by the following steps:

[0007] S1: Collect data on the target wind farm and nearby wind farms for training and testing of the prediction model;

[0008] S2: Train the discriminator of W-GAN (Wasserstein GAN) with real data, generate data using the generator, and use the results of the discriminator to train the generator; train the prediction model of neighboring wind farms and target wind farms based on support vector regression (SVR), and analyze the error distribution of itself and neighboring wind farms; then analyze the correlation of neighboring wind farms based on the maximum mutual information coefficient analysis method, and classify the neighboring wind farm data according to the strength of the correlation to form a dataset to be transferred, which together with the W-GAN generated dataset constitutes the target wind farm training dataset.

[0009] S3: The large amount of data generated by the trained W-GAN is combined with the data of neighboring wind farms with similar error distributions to be transferred and predicted by the CNN-LSTM combined model. The prediction results are obtained by combining a large amount of information in the generated data and the time series information in the actual training data. Then, the discrete wind power probability prediction results at the corresponding time are obtained through the quantile regression layer. Then, the corresponding continuous probability density expression is obtained through the kernel density estimation method. The probability prediction results are obtained through the point prediction error distribution statistics.

[0010] S4: Using ACE (Average Coverage Error), AW (Average Width), and WINKLER as performance evaluation criteria for wind power probabilistic prediction models, the most suitable CNN-LSTM combined probabilistic prediction model is selected. The test data from S1 is input into the selected CNN-LSTM combined probabilistic prediction model for prediction, and the results are output.

[0011] Furthermore, step S1 specifically includes the following steps:

[0012] S1.1: Collect data from the target wind farm with a time resolution of 15 minutes for 40 consecutive days, and data from neighboring wind farms for 1 consecutive year, including numerical weather prediction (NWP) data and their corresponding power data. Divide 75% of the collected data from the target wind farm into a training dataset and 25% into a test dataset, which will be used for training and testing the prediction model, respectively. Use 75% of the collected data from neighboring wind farms for model training.

[0013] Further, step S2 specifically includes: training the discriminator of W-GAN using a training dataset; training the generator and generating new data using the discriminator results; generating data using random noise using the generator; testing the discriminator using real training datasets and generated data; and feeding back the test results to train the generator and discriminator. The above training process is repeated until Nash equilibrium is reached, which is the validity test of the generated data based on principal component analysis, and the generated data constitutes the W-GAN generated dataset.

[0014] S2.2: First, the training dataset is used to train the target wind farm prediction model based on SVR, and 75% of the data from neighboring wind farms is used to train the neighboring wind farm prediction model based on SVR. The error distribution of the target wind farm prediction model and the neighboring wind farm prediction model is statistically analyzed. Then, the correlation of error distributions of each wind farm is analyzed by the error distribution correlation analysis method of maximum mutual information coefficient (MIC). The neighboring wind farm data are classified according to the strength of the correlation. The relevant data of the target wind farm with similar error distribution are selected to construct the transfer dataset. The transfer dataset and the W-GAN generated dataset together constitute the target wind farm training dataset.

[0015] Furthermore, step S3 specifically includes the following steps:

[0016] S3.1: Input the W-GAN generated dataset trained in step S2.1 into the CNN model. The CNN model extracts a large number of spectral features from the generated dataset. Input the training dataset into the LSTM model. The LSTM model extracts its own time series features from the actual training dataset. Input the dataset to be transferred trained in step S2.2 into the LSTM model. The LSTM model extracts the time series features of the dataset to be transferred from the neighboring wind farms. Use the dataset to be transferred to fine-tune the transfer learning model. Then, fuse the features obtained from the above three models through the fully connected layer of the CNN-LSTM combined probabilistic prediction model to obtain the fused features.

[0017] S3.2: Input the fused features obtained in step S3.1 into the CNN-LSTM combined probability prediction model, use the quantile regression layer of the Pinball loss function for prediction, and obtain the corresponding discretized wind power probability prediction results under different confidence intervals;

[0018] S3.3: Input the discretized probability prediction results obtained in step S3.2 into the kernel density estimation model, and select the uniform kernel function, triangular kernel function and Gaussian kernel function to obtain the corresponding continuous probability density expression;

[0019] S3.4: By inputting the fused features obtained in step S3.1 into the CNN-LSTM combined probabilistic prediction model, the regression layer using the mean square error function is used to obtain the point prediction results and the statistical error distribution is used to obtain the probabilistic prediction results. The discretized wind power probabilistic prediction results, the continuous probability density expression, and the probabilistic prediction results together constitute the probabilistic prediction expression.

[0020] Furthermore, in step S4, the specific steps include: using ACE, AW, and WINKLER indices as performance evaluation criteria for wind power probability prediction models, selecting the CNN-LSTM combined model with the optimal parameters as the short-term wind power probability prediction model, and using the test dataset to select the most suitable short-term wind power probability prediction model for prediction, and outputting the results.

[0021] The beneficial effects achieved by this invention are as follows: The method of this invention expands the sample pool in scenarios with insufficient samples, such as newly built wind farms and extreme weather, by using an improved Generative Adversarial Network (Wasserstein GAN) and data transfer learning. Furthermore, in the prediction modeling stage, CNN and LSTM are used to learn the information contained in the generated and original samples respectively, improving the final prediction accuracy and playing a crucial role in the probability prediction of wind power in scenarios with insufficient samples. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process structure of the present invention. Detailed Implementation

[0023] The following is a detailed description of the embodiments of the present invention with reference to the accompanying drawings. The purpose of the present invention is to provide a wind power probability prediction method based on data augmentation and CNN-LSTM.

[0024] like Figure 1 As shown, a wind power probability prediction method based on data augmentation and CNN-LSTM is characterized by the following steps:

[0025] S1: Collect data from the target wind farm and neighboring wind farms for training and testing the prediction model. Specific steps include: collecting data from the target wind farm over 40 consecutive days with a time resolution of 15 minutes, and data from neighboring wind farms over one year, including NWP data and its corresponding power data; dividing 75% of the collected data from the target wind farm into a training dataset and 25% into a test dataset, respectively for training and testing the prediction model; and using 75% of the collected data from neighboring wind farms for model training.

[0026] S2: Train the discriminator of W-GAN (Wasserstein GAN) using real data, generate data using the generator, and use the discriminator's results to train the generator; train prediction models for neighboring wind farms and the target wind farm based on support vector regression (SVR), and analyze the error distribution of the model and neighboring wind farms; then analyze the correlation of neighboring wind farms using the maximum mutual information coefficient analysis method, and classify the neighboring wind farm data according to the strength of the correlation to form a dataset to be transferred, which together with the W-GAN generated dataset constitutes the target wind farm training dataset; the specific steps include:

[0027] S2.1: Train the discriminator of W-GAN using the training dataset, use the discriminator results to train the generator and generate new generated data, and then use the generator to generate data through random noise; test the discriminator using the real training dataset and the generated data, and use the test results to train the generator and discriminator; repeat the above training process until Nash equilibrium is reached, which is the validity test of the generated data based on principal component analysis, and then use the generated data to form the W-GAN generated dataset;

[0028] S2.2: First, the training dataset is used to train the target wind farm prediction model based on SVR, and 75% of the data from neighboring wind farms is used to train the neighboring wind farm prediction model based on SVR. The error distribution of the target wind farm prediction model and the neighboring wind farm prediction model is statistically analyzed. Then, the correlation of error distribution of each wind farm is analyzed by the error distribution correlation analysis method of maximum mutual information coefficient (MIC). The neighboring wind farm data is classified according to the strength of the correlation. The relevant data of the target wind farm with similar error distribution is selected to construct the transfer dataset. The transfer dataset and the W-GAN generated dataset together constitute the target wind farm training dataset.

[0029] S3: The large amount of data generated by the trained W-GAN is combined with the data from neighboring wind farms with similar error distributions to be transferred, and then predicted using a CNN-LSTM combined model. This yields prediction results that combine a large amount of information from the generated data with the temporal information from the actual training data. Then, a quantile regression layer is used to obtain the discrete wind power probability prediction results for the corresponding time points. Finally, a kernel density estimation method is used to obtain the corresponding continuous probability density expression, and the probability prediction results are obtained through point prediction error distribution statistics. The specific steps include:

[0030] S3.1: Input the W-GAN generated dataset trained in step S2.1 into the CNN model. The CNN model extracts a large number of spectral features from the generated dataset. Input the training dataset into the LSTM model. The LSTM model extracts its own time series features from the actual training dataset. Input the dataset to be transferred trained in step S2.2 into the LSTM model. The LSTM model extracts the time series features of the dataset to be transferred from the neighboring wind farms. Use the dataset to be transferred to fine-tune the transfer learning model. Then, fuse the features obtained from the above three models through the fully connected layer of the CNN-LSTM combined probabilistic prediction model to obtain the fused features.

[0031] S3.2: Input the fused features obtained in step S3.1 into the CNN-LSTM combined probability prediction model, use the quantile regression layer of the Pinball loss function for prediction, and obtain the corresponding discretized wind power probability prediction results under different confidence intervals;

[0032] S3.3: Input the discretized probability prediction results obtained in step S3.2 into the kernel density estimation model, and select the uniform kernel function, triangular kernel function and Gaussian kernel function to obtain the corresponding continuous probability density expression;

[0033] S3.4: By inputting the fused features obtained in step S3.1 into the CNN-LSTM combined probabilistic prediction model, the regression layer using the mean square error function is used to obtain the point prediction results and the statistical error distribution is used to obtain the probabilistic prediction results. The discretized wind power probabilistic prediction results, the continuous probability density expression, and the probabilistic prediction results together constitute the probabilistic prediction expression.

[0034] S4: Using ACE (Average Coverage Error), AW (Average Width), and WINKLER as performance evaluation criteria for wind power probability prediction models, the CNN-LSTM combined model with the best parameters is selected as the short-term wind power probability prediction model, and the test dataset is used to select the most suitable short-term wind power probability prediction model for prediction, and the results are output.

Claims

1. A wind power probability prediction method based on data augmentation and CNN-LSTM, characterized in that... Follow these steps: S1: Collect data from the target wind farm and neighboring wind farms for training and testing the prediction model; divide 75% of the collected data from the target wind farm into a training dataset; divide 25% of the collected data from the target wind farm into a test dataset. S2: Train the discriminator of W-GAN using real data, generate data using the generator, and use the results of the discriminator to train the generator, thus forming the W-GAN generated dataset; train the prediction model of neighboring wind farms and target wind farms based on support vector regression (SVR) to analyze the error distribution of itself and neighboring wind farms; then analyze the correlation of neighboring wind farms based on the maximum mutual information coefficient analysis method, and classify the neighboring wind farm data according to the strength of the correlation to form the dataset to be transferred, which together with the W-GAN generated dataset constitutes the target wind farm training dataset. S3: Combine the large amount of data generated by the trained W-GAN with the data of neighboring wind farms with similar error distributions to be transferred, and use the CNN-LSTM combined model for prediction to obtain the prediction results that combine a large amount of information in the generated data with the temporal information in the actual training data; specifically: input the W-GAN generated dataset into the CNN model to extract a large number of spectral features; The training dataset is input into the LSTM model to extract its own time-series features. The dataset to be transferred is input into the LSTM model to extract the time-series features of the dataset to be transferred from neighboring wind farms. The transfer learning model is then fine-tuned using the dataset to be transferred. The features obtained from the above three models are then fused through the fully connected layer of the CNN-LSTM combined probabilistic prediction model to obtain fused features. The fused features are then input into the CNN-LSTM combined probabilistic prediction model to obtain discretized wind power probability prediction results. The kernel density estimation method is then used to obtain the corresponding continuous probability density expression. The probability prediction result is obtained by statistically analyzing the point prediction error distribution. The discretized wind power probability prediction results, the continuous probability density expression, and the probability prediction result obtained by statistically analyzing the point prediction error distribution together constitute the probability prediction expression. S4: Using ACE, AW, and WINKLER indices as performance evaluation criteria for wind power probabilistic prediction models, the most suitable CNN-LSTM combined probabilistic prediction model is selected. The test data from the S1 test dataset is input into the selected CNN-LSTM combined probabilistic prediction model for prediction, and the results are output.

2. The wind power probability prediction method based on data augmentation and CNN-LSTM as described in claim 1, characterized in that, In step S1, the specific steps are as follows: collect data from the target wind farm with a time resolution of 15 minutes for 40 consecutive days, and data from neighboring wind farms for 1 consecutive year, including NWP data and its corresponding power data, which are used for training and testing the prediction model respectively; and use 75% of the data collected from neighboring wind farms for model training.

3. The wind power probability prediction method based on data augmentation and CNN-LSTM according to claim 2, characterized in that, The specific steps in step S2 include: S2.1: Train the discriminator of W-GAN using the training dataset, use the discriminator results to train the generator and generate new generated data, and then use the generator to generate data through random noise; test the discriminator using the real training dataset and the generated data, and use the test results to train the generator and discriminator; repeat the above training process until Nash equilibrium is reached, which is the validity test of the generated data based on principal component analysis, and then use the generated data to form the W-GAN generated dataset; S2.2: First, the training dataset is used to train the target wind farm prediction model based on SVR, and 75% of the data from neighboring wind farms is used to train the neighboring wind farm prediction model based on SVR. The error distribution of the target wind farm prediction model and the neighboring wind farm prediction model is statistically analyzed. Then, the correlation of error distribution of each wind farm is analyzed by the error distribution correlation analysis method of maximum mutual information coefficient analysis. The neighboring wind farm data is classified according to the strength of the correlation. The relevant data of the target wind farm with similar error distribution is selected to construct the transfer dataset. The transfer dataset and the W-GAN generated dataset together constitute the target wind farm training dataset.

4. The wind power probability prediction method based on data augmentation and CNN-LSTM as described in claim 3, characterized in that, The specific steps in step S3 include: S3.1: Input the W-GAN generated dataset trained in step S2.1 into the CNN model, the training dataset into the LSTM model, and the dataset to be transferred trained in S2.2 into the LSTM model to obtain a large number of spectral features of the W-GAN generated dataset, the time series features of the training dataset, and the time series features of the dataset to be transferred, respectively, and fine-tune the transfer learning model; then fuse the features obtained from the above three models through the fully connected layer of the CNN-LSTM combined probabilistic prediction model to obtain the fused features; S3.2: Input the fused features obtained in step S3.1 into the CNN-LSTM combined probability prediction model, use the quantile regression layer of the Pinball loss function for prediction, and obtain the corresponding discretized wind power probability prediction results under different confidence intervals; S3.3: Input the discretized probability prediction results obtained in step S3.2 into the kernel density estimation model, and select the uniform kernel function, triangular kernel function and Gaussian kernel function to obtain the corresponding continuous probability density expression; S3.4: By inputting the fused features obtained in step S3.1 into the CNN-LSTM combined probabilistic prediction model, the regression layer using the mean square error function is used to obtain the point prediction results and the statistical error distribution is used to obtain the probabilistic prediction results. The discretized wind power probabilistic prediction results, the continuous probability density expression, and the probabilistic prediction results obtained by statistically analyzing the point prediction error distribution together constitute the probabilistic prediction expression.

5. The wind power probability prediction method based on data augmentation and CNN-LSTM as described in claim 4, characterized in that, The specific steps in step S4 are as follows: Using ACE, AW, and WINKLER indices as performance evaluation criteria for wind power probability prediction models, the CNN-LSTM combined model with the optimal parameters is selected as the short-term wind power probability prediction model. The test dataset is used to select the most suitable short-term wind power probability prediction model for prediction, and the results are output.

Images