Monthly-scale wind power prediction method based on WICNN and TPA-LSTM
Through the data processing method combined with WICNN and the TPA-LSTM model, the problem of insufficient time-sequence relationship and space-time coupling characteristic capture in monthly-scale wind power prediction is solved, and a prediction effect with higher accuracy and robustness is achieved.
Patent Information
- Application Number
- CN202510594796.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2025-04-15
- Filing Date
- 2025-05-09
- Publication Date
- 2025-08-08
AI Technical Summary
Traditional wind power power prediction methods are difficult to capture the complex timing relationships and cross-scale space-time coupling characteristics over long-term spans on the monthly scale, and ignore the impact of wind farm spatial heterogeneity and climate system changes, resulting in insufficient prediction accuracy and generalization capabilities.
The wavelet decomposition and adaptive filtering network framework are used to denoise data, combined with the wavelet integrated convolutional neural network WICNN to extract time-frequency features, and a prediction model combining the time-mode attention mechanism TPA and LSTM is introduced to optimize the hidden layer and capture important patterns and features of the monthly-scale wind power time series.
It significantly improves the accuracy and robustness of monthly-scale wind power power prediction, can effectively capture non-stationary characteristics and long-term dependencies, and improves the accuracy and reliability of prediction.
Smart Images
Figure CN120454049A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of new energy power prediction and relates to a monthly-scale wind power prediction method based on WICNN and TPA-LSTM. Background Art
[0002] As installed wind power capacity increases, the randomness and volatility of wind power generation pose a significant challenge to the stable operation of power systems. Accurate monthly wind power forecasting is one effective way to address this challenge. Monthly forecasts are also beneficial for guiding long-term power system scheduling and regular maintenance of wind farms.
[0003] Monthly wind power forecasting involves not only analyzing medium- and long-term trends in wind power output sequences but also fully accounting for complex nonlinear fluctuations caused by seasonal variations and significant weather disturbances. At this time scale, wind power exhibits stronger time dependence and potential cyclical variations, placing higher demands on the time series modeling capabilities of forecasting models. Therefore, it is necessary to establish a modeling framework that can effectively extract long-term time series structure and capture non-stationary characteristics to improve forecast accuracy and robustness. Traditional wind power forecasting methods are generally more suitable for short-term or daily forecasts. When applied to monthly wind power forecasting, these methods often face insufficient model capabilities, making it difficult to capture complex time series relationships over long time spans and cross-scale spatiotemporal coupling characteristics. Furthermore, traditional methods generally ignore the spatial heterogeneity of wind farm locations and the potential impact of climate system changes on wind energy resources, resulting in limited forecast accuracy and weak generalization. Summary of the Invention
[0004] The purpose of the present invention is to solve the problems existing in the above-mentioned background technology and provide a monthly scale wind power forecasting method based on WICNN and TPA-LSTM. In order to solve the significant quality fluctuation problem of input meteorological data and historical power data, a dynamic wavelet decomposition and adaptive filtering network framework is constructed, and a multi-scale adaptive threshold selection module (ATSM) is developed. The noise reduction parameter configuration is dynamically optimized according to the time-frequency characteristics of each decomposition level, significantly improving the input data quality; then the previous step data is input into the wavelet integrated convolutional neural network WICNN, which designs a dual-domain feature interaction fusion mechanism and establishes a cross-domain attention gating unit. The dynamic weight allocation of the feature tensor is realized through the time-frequency dual-channel convolution operation, effectively extracting the cross-domain deep spatiotemporal coupling features, laying the foundation for subsequent prediction; finally, the traditional attention mechanism is difficult to effectively learn the long-term dependency of wind power time series data at the monthly scale. The present invention proposes a prediction model combining the temporal pattern attention mechanism (TPA) and the long short-term memory neural network (LSTM); the extracted feature information is input into the TPA-LSTM to fully capture the important patterns and features in the monthly time series, effectively improving the monthly scale wind power forecast accuracy.
[0005] The purpose of the present invention is achieved by the following technical measures:
[0006] A monthly wind power forecasting method based on WICNN and TPA-LSTM includes the following steps:
[0007] S1: Perform wavelet decomposition on the original wind farm data and perform denoising processing based on adaptively selected thresholds according to different frequencies to improve data quality;
[0008] S2: The data processed in step S1 is input into the wavelet integrated convolutional neural network WICNN, and the time domain and frequency domain features of the data are extracted through dual-channel convolution operations in the time domain and frequency domain;
[0009] S3: The temporal pattern attention mechanism (TPA) is introduced to optimize the LSTM hidden layer. TPA performs a convolution operation on the monthly wind power time series in the LSTM hidden layer and updates the LSTM hidden layer. TPA is combined with the LSTM that updates the LSTM hidden layer to obtain the TPA-LSTM prediction model.
[0010] S4: The time domain and frequency domain features extracted in step S2 are input into the TPA-LSTM prediction model for training. After training, the final TPA-LSTM prediction model is obtained. The numerical weather forecast data of the wind farm for the next month is input into the final TPA-LSTM prediction model for prediction to obtain the final monthly wind power prediction result.
[0011] Furthermore, the specific steps of S1 include:
[0012] S1.1: Select Daubechies6 wavelet basis to perform wavelet decomposition on the original wind farm power data to obtain five decomposition signal frequency bands;
[0013] S1.2: Calculate the noise variance and signal-to-noise ratio for the five decomposed signal frequency segments obtained in S1.1.
[0014] S1.3: Evaluate the denoising threshold based on the calculation results and perform denoising;
[0015] S1.4: Perform wavelet reconstruction on the denoised data in S1.3 to obtain the noise-filtered data.
[0016] Furthermore, the specific steps of S2 include:
[0017] S2.1: Duplicate the filtered data into three copies for subsequent convolution operations, with one copy remaining unchanged.
[0018] S2.2: The second part performs 1×1 and 3×3 convolution operations to extract time domain feature information;
[0019] S2.3: After the third part performs a 1×1 convolution operation, it performs a wavelet decomposition and a 3×3 convolution operation on the frequency domain data to extract the frequency domain feature information;
[0020] S2.4: Concatenate and fuse the three data sets to obtain data with time domain and frequency domain feature information.
[0021] Furthermore, the specific steps of S3 include:
[0022] S3.1: Construct LSTM and determine the model parameters of hidden layer size, learning rate, loss function, and optimizer; link the LSTM hidden layer to the temporal pattern attention mechanism (TPA);
[0023] S3.2: Perform convolution operations on the hidden layer monthly wind power time series through TPA’s 1D CNN to generate an information matrix;
[0024] S3.3: Input the obtained information matrix into the scoring function and compare it with the current hidden state to obtain the information matrix weight;
[0025] S3.4: Normalize the information matrix weights and perform weighted summation of the information matrix according to the weights to obtain a new information matrix containing monthly scale information;
[0026] S3.5: Concatenate the new information matrix containing the monthly scale information obtained in S3.4 with the current hidden state of the LSTM, update the current hidden state information of the LSTM, combine the LSTM with the updated hidden layer with TPA, and obtain the TPA-LSTM prediction model.
[0027] Furthermore, the specific steps of S4 include:
[0028] S4.1: Input the time domain and frequency domain feature information sequences extracted in S2.4 into the TPA-LSTM prediction model for training. Stop training after the validation set loss does not decrease after 10 iterations to obtain the final TPA-LSTM prediction model.
[0029] S4.2: Input the numerical weather forecast data of the wind farm for the next month into the final TPA-LSTM prediction model to predict and obtain the final monthly wind power.
[0030] The beneficial effects achieved by the present invention are as follows: first, the present invention constructs an adaptive threshold selection module to dynamically adjust and denoise wavelet decomposition signals with different frequencies, noise variances, and signal-to-noise ratios, thereby effectively improving data quality, retaining the original data information to the greatest extent, and reducing the impact of noise on the prediction model; then the denoised data is placed in the time domain and frequency domain dual-domain feature interactive fusion module designed by the present invention, which can analyze the wind power change trend and potential changes from two perspectives, and more comprehensively capture non-stationary and nonlinear time series characteristics; finally, the time attention mechanism TPA is introduced, which can optimize the LSTM hidden layer's ability to extract time information at the monthly scale, and can also use attention weights to focus on important time patterns related to the current time at the monthly scale, fully tap the global features during medium- and long-term power forecasting, and improve the accuracy and reliability of forecasts at the monthly scale. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 It is a structural block diagram of the overall process of the present invention. DETAILED DESCRIPTION
[0032] The following is a detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. The purpose of the present invention is to provide a monthly wind power prediction method based on WICNN and TPA-LSTM.
[0033] like Figure 1 As shown in FIG, a monthly wind power forecasting method based on WICNN and TPA-LSTM includes the following steps:
[0034] S1: Perform wavelet decomposition on the original wind farm data and perform denoising based on adaptively selected thresholds at different frequencies to improve data quality. The specific steps are as follows:
[0035] S1.1: Select Daubechies6 wavelet basis to perform wavelet decomposition on the original wind farm power data to obtain five decomposition signal frequency bands;
[0036] S1.2: Calculate the noise variance and signal-to-noise ratio for the five decomposed signal frequency segments obtained in S1.1.
[0037] S1.3: Evaluate the denoising threshold based on the calculation results and perform denoising;
[0038] S1.4: Perform wavelet reconstruction on the denoised data in S1.3 to obtain the noise-filtered data.
[0039] S2: Input the data processed in step S1.3 into the wavelet integrated convolutional neural network WICNN, and extract the time domain and frequency domain features of the data through dual-channel convolution operations in the time domain and frequency domain. The specific steps are as follows:
[0040] S2.1: Duplicate the filtered data into three copies for subsequent convolution operations, with one copy remaining unchanged.
[0041] S2.2: The second part performs 1×1 and 3×3 convolution operations to extract time domain feature information;
[0042] S2.3: After the third part performs a 1×1 convolution operation, it performs a wavelet decomposition and a 3×3 convolution operation on the frequency domain data to extract the frequency domain feature information;
[0043] S2.4: Concatenate and fuse the three data sets to obtain data with time domain and frequency domain feature information.
[0044] S3: Introduce the temporal pattern attention mechanism (TPA) to optimize the LSTM hidden layer. TPA performs a convolution operation on the monthly wind power time series in the LSTM hidden layer and updates the LSTM hidden layer. TPA is combined with the LSTM in the updated LSTM hidden layer to obtain the TPA-LSTM prediction model. The specific steps are as follows:
[0045] S3.1: Construct LSTM and determine the model parameters of hidden layer size, learning rate, loss function, and optimizer; link the LSTM hidden layer to the temporal pattern attention mechanism (TPA);
[0046] S3.2: Perform convolution operations on the hidden layer monthly wind power time series through TPA’s 1D CNN to generate an information matrix;
[0047] S3.3: Input the obtained information matrix into the scoring function and compare it with the current hidden state to obtain the information matrix weight;
[0048] S3.4: Normalize the information matrix weights and perform weighted summation of the information matrix according to the weights to obtain a new information matrix containing monthly scale information;
[0049] S3.5: Concatenate the new information matrix containing the monthly scale information obtained in S3.4 with the current hidden state of the LSTM, update the current hidden state information of the LSTM, combine the LSTM with the updated hidden layer with TPA, and obtain the TPA-LSTM prediction model.
[0050] S4: Input the time domain and frequency domain features extracted in step S2.4 into the TPA-LSTM prediction model for training. After training, the final TPA-LSTM prediction model is obtained. The numerical weather forecast data of the wind farm for the next month is input into the final TPA-LSTM prediction model for prediction to obtain the final monthly wind power prediction result. The specific steps are as follows:
[0051] S4.1: Input the time domain and frequency domain feature information sequences extracted in S2.4 into the TPA-LSTM prediction model for training. Stop training after the validation set loss does not decrease after 10 iterations to obtain the final TPA-LSTM prediction model.
[0052] S4.2: Input the numerical weather forecast data of the wind farm for the next month into the final TPA-LSTM prediction model to predict and obtain the final monthly wind power.
Claims
1. A monthly wind power forecasting method based on WICNN and TPA-LSTM, characterized by Follow these steps: S1: Perform wavelet decomposition on the original wind farm data and perform denoising processing based on adaptively selected thresholds according to different frequencies to improve data quality; S2: The data processed in step S1 is input into the wavelet integrated convolutional neural network WICNN, and the time domain and frequency domain features of the data are extracted through dual-channel convolution operations in the time domain and frequency domain; S3: The temporal pattern attention mechanism (TPA) is introduced to optimize the LSTM hidden layer. TPA performs a convolution operation on the monthly wind power time series in the LSTM hidden layer and updates the LSTM hidden layer. TPA is combined with the LSTM that updates the LSTM hidden layer to obtain the TPA-LSTM prediction model. S4: The time domain and frequency domain features extracted in step S2 are input into the TPA-LSTM prediction model for training. After training, the final TPA-LSTM prediction model is obtained. The numerical weather forecast data of the wind farm for the next month is input into the final TPA-LSTM prediction model for prediction to obtain the final monthly wind power prediction result.
2. The monthly wind power forecasting method based on WICNN and TPA-LSTM according to claim 1 is characterized in that: In the step S1, the specific steps include: S1.1: Select Daubechies6 wavelet basis to perform wavelet decomposition on the original wind farm power data to obtain five decomposition signal frequency bands; S1.2: Calculate the noise variance and signal-to-noise ratio for the five decomposed signal frequency segments obtained in S1.
1. S1.3: Evaluate the denoising threshold based on the calculation results and perform denoising; S1.4: Perform wavelet reconstruction on the denoised data in S1.3 to obtain the noise-filtered data.
3. The monthly wind power forecasting method based on WICNN and TPA-LSTM according to claim 2 is characterized in that: In the step S2, the specific steps include: S2.1: Duplicate the filtered data into three copies for subsequent convolution operations, with one copy remaining unchanged. S2.2: The second part performs 1×1 and 3×3 convolution operations to extract time domain feature information; S2.3: After the third part performs a 1×1 convolution operation, it performs a wavelet decomposition and a 3×3 convolution operation on the frequency domain data to extract the frequency domain feature information; S2.4: Concatenate and fuse the three data sets to obtain data with time domain and frequency domain feature information.
4. The monthly wind power forecasting method based on WICNN and TPA-LSTM according to claim 3 is characterized in that: In the step S3, the specific steps include: S3.1: Construct LSTM and determine the model parameters of hidden layer size, learning rate, loss function, and optimizer; link the LSTM hidden layer to the temporal pattern attention mechanism (TPA); S3.2: Perform convolution operations on the hidden layer monthly wind power time series through TPA’s 1D CNN to generate an information matrix; S3.3: Input the obtained information matrix into the scoring function and compare it with the current hidden state to obtain the information matrix weight; S3.4: Normalize the information matrix weights and perform weighted summation of the information matrix according to the weights to obtain a new information matrix containing monthly scale information; S3.5: Concatenate the new information matrix containing the monthly scale information obtained in S3.4 with the current hidden state of the LSTM, update the current hidden state information of the LSTM, combine the LSTM with the updated hidden layer with TPA, and obtain the TPA-LSTM prediction model.
5. The monthly wind power forecasting method based on WICNN and TPA-LSTM according to claim 4 is characterized in that: In the step S4, the specific steps include: S4.1: Input the time domain and frequency domain feature information sequences extracted in S2.4 into the TPA-LSTM prediction model for training. Stop training after the validation set loss does not decrease after 10 iterations to obtain the final TPA-LSTM prediction model. S4.2: Input the numerical weather forecast data of the wind farm for the next month into the final TPA-LSTM prediction model to predict and obtain the final monthly wind power.