A wind speed multi-scale cascade enhanced end-to-end prediction method

Abstract

The application discloses a kind of wind speed multiscale cascade enhancement end-to-end prediction method, comprising the following steps: S1. wind speed data is obtained;S2. wind speed data obtained in step S1 is preprocessed, and wind speed data set is obtained;S3. based on multiscale prediction strategy, initial wind speed prediction model is constructed;S4. wind speed data set obtained in step S2 is used, and initial wind speed prediction model obtained in step S3 is trained, and wind speed prediction model is obtained;S5. wind speed prediction model obtained in step S4 is used, and actual wind speed prediction is carried out.The method of the application uses the convolutional neural network structure design based on deep learning, realizes the accurate prediction of multiscale wind speed, has high accuracy, low false detection rate, fast detection speed, good real-time performance, effectively improves the prediction efficiency, and has broad application prospect.

Description

Technical Field

[0001] This invention belongs to the field of wind speed prediction, specifically relating to a multi-scale cascaded enhancement end-to-end prediction method for wind speed. Background Technology

[0002] Currently, with increasing concerns about energy sustainability and scarcity, the energy sector is actively developing renewable energy resources and combining energy production and use to further conserve energy. Accurately quantifying production capacity indicators is crucial for energy development and energy policy formulation.

[0003] In the field of new energy power generation, wind power plays a vital role. The principle of wind power generation is to convert the kinetic energy of wind into electrical energy; therefore, wind energy is renewable and inexhaustible. At the same time, wind power generation is very clean and has good environmental benefits. However, as an energy type with strong randomness and seasonality, the intermittency and instability of wind power production pose significant challenges to its effective utilization.

[0004] Therefore, wind speed forecasting is of great significance for the safe operation of the power grid and the assurance of power quality. Currently, the main methods for wind speed forecasting include methods based on numerical weather prediction, statistical methods based on existing wind speed data, and machine learning methods. Among these, machine learning methods have shown great potential since their emergence, capable of handling nonlinear relationships and complex patterns, with a higher accuracy ceiling than traditional methods. The accuracy of wind speed forecasting plays a decisive role in wind power forecasting and is of great importance to the operation of wind farms and power systems. However, current methods for forecasting wind speed using machine learning are not yet mature enough to provide long-term, accurate, and effective wind speed forecasts. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-scale cascaded enhancement end-to-end prediction method for wind speed, which can improve the accuracy and length of wind speed prediction and effectively enhance prediction precision.

[0006] This invention provides a multi-scale cascaded enhancement end-to-end prediction method for wind speed, comprising the following steps:

[0007] S1. Obtain wind speed data;

[0008] S2. Preprocess the wind speed data obtained in step S1 to obtain a wind speed dataset;

[0009] S3. Construct an initial wind speed prediction model based on a multi-scale prediction strategy;

[0010] S4. Using the wind speed dataset obtained in step S2, train the initial wind speed prediction model obtained in step S3 to obtain the wind speed prediction model.

[0011] S5. Use the wind speed prediction model obtained in step S4 to make actual wind speed predictions.

[0012] The wind speed data mentioned in step S1 is obtained through wind speed sensors at ground monitoring points, including temperature parameters, air pressure parameters, and wind speed parameters.

[0013] Step S2 specifically involves: First, preprocessing the wind speed data obtained in step S1, constructing the preprocessed wind speed data detected by N stations into a 1*L*W format, where L is the long-term wind speed data recording speed, and W is the number of multi-source parameters obtained. Annotations are added to each station where the wind speed data was acquired based on the station's latitude and longitude relationship, filling in missing data with the value "0". A prediction parameter λ is added to the wind speed data: the short-term prediction parameter range is 0 < λ ≤ 96; the medium-term prediction parameter range is 96 < λ ≤ 192; and the long-term prediction parameter range is λ > 192.

[0014] The preprocessing includes data missing completion, data anomaly handling, and data standardization. Regarding data missing, forward imputation is used to complete the missing data based on the discontinuous nature of the missing data. Regarding data anomalies, a probabilistic autoencoder-based method is selected to process the data. Regarding data standardization, Z-score standardization is used to process the data, taking into account the inconsistency between wind speed and wind vector dimensions and the requirements of the training model.

[0015] In step S3, the initial wind speed prediction model includes a first RevIN module, a judgment module, a short-term prediction branch module, a medium-term prediction branch module, a long-term prediction branch module, a feature fusion module, and a second RevIN module.

[0016] The data input to the model is first normalized by the first RevIN module before entering the judgment module; the judgment module makes the following judgments based on the predicted judgment parameters:

[0017] If the prediction judgment parameter 0 < λ ≤ 96, the output of the judgment module enters the short-term prediction branch module; after being processed by the short-term prediction module, it is then fused through the feature fusion module, and the output result is inversely normalized through the second RevIN module to obtain the model output result.

[0018] If the prediction judgment parameter 96 < λ ≤ 192, the output of the judgment module enters the intermediate prediction branch module; after being processed by the intermediate prediction module, the feature fusion module performs feature fusion, and the output result is inversely normalized by the second RevIN module to obtain the model output result.

[0019] If the prediction judgment parameter λ > 192, the output of the judgment module enters the long-term prediction branch module; after being processed by the long-term prediction module, it is then fused through the feature fusion module, and the output result is inversely normalized through the second RevIN module to obtain the model output result.

[0020] The first RevIN module normalizes the input data by calculating the mean and standard deviation, expressed using the following formula:

[0021]

[0022] in, L represents the average value of the input data. p The length of the input data; For each data value; j represents data from different dimensions at the same time, such as dimensions including temperature and air pressure; The deviation of the input data; after obtaining the mean and standard deviation of the data, the input data is normalized using the following formula:

[0023]

[0024] in, The input data is normalized; ε is the disturbance error; the normalized input data is input into the judgment module.

[0025] The short-term prediction branch modules include the MCIE-C branch module, the MCIE-L branch module, and the MCIE-T branch module;

[0026] The intermediate-term prediction branch module includes an MCIE-T branch module, a first MCIE-L branch module, and a second MCIE-L branch module; the first MCIE-L branch module and the second MCIE-L branch module have the same structure.

[0027] The long-term prediction branch module includes a first MCIE-T branch module, a second MCIE-T branch module, and a third MCIE-T branch module; the first to third MCIE-T branch modules have the same structure.

[0028] The MCIE-C branch module includes a first convolutional coding module, a first 1×1 convolutional layer module, a 3×3 depthwise separable convolutional layer module, a concat module, a second 1×1 convolutional layer module, and a second convolutional coding module, which are connected in sequence.

[0029] The MCIE-C branch module performs dimensionality transformation on the input data through the first convolutional coding module, obtains local features through the first 1×1 convolutional layer module and the 3×3 depthwise separable convolutional layer module, and finally exchanges dimensionality information through the second 1×1 convolutional layer module to integrate fine-grained features; the output of the second 1×1 convolutional layer module is processed by the second convolutional coding module and used as the output of the MCIE-C branch module.

[0030] The first 1×1 convolutional layer module has the same structure as the second 1×1 convolutional layer module, including a 1×1 convolutional layer, a ReLU activation function, and a post-activation BatchNorm connected in series.

[0031] The first convolutional coding module performs dimensionality transformation on the input data. It uses a sliding window approach, moving along the time dimension of the input data according to the window length and stride, and extracting subsequence patches as the output. This is expressed by the following formula:

[0032]

[0033] Where P is the window length; S is the step size; and D is the data dimension.

[0034] A 3×3 depthwise separable convolutional layer module captures the global receptive field and local location features in the input data. The entire process inputs N / 2 patches, and each patch undergoes a separate convolution operation with a kernel to generate N feature maps. The size of the convolution kernel is the same as the patch stride in the first convolutional encoding module, as expressed by the following formula:

[0035]

[0036] Where l is the length of the input data;

[0037] The concat module takes the output of the first 1×1 convolutional layer module and the output of the 3×3 depth-separable convolutional layer module as inputs, and concatenates them to create an input with N channels;

[0038] The MCIE-L branch module separates linear and nonlinear features and extracts them separately. The MCIE-L branch module includes a Fourier module, a linear feature extraction module, and a nonlinear feature extraction module. The input data is first processed by the Fourier module to decompose it into linear and nonlinear data. The linear feature extraction module uses the linear data output from the Fourier module as input; the nonlinear feature extraction module uses the nonlinear data output from the Fourier module as input; the output of the MCIE-L branch module is the direct sum of the outputs from the linear and nonlinear feature extraction modules.

[0039] The Fourier module clarifies the sequence components of the input data, calculates and sorts the average amplitude of different data in the frequency domain using the Fast Fourier Transform technique for each past window, and decomposes the input data into linear and nonlinear data, as shown in the following formula:

[0040] X L =FTT -1 (X in )

[0041] X NonL =X in -FTT -1 (X in )

[0042] Among them, X L Linear data; X NonL For non-linear data; X in For the input data of the Fourier module; FTT -1 Inverse Fourier transform;

[0043] The linear feature extraction module comprises a PatchPartition module, a first Tanh activation function, a first MLP module, a Least-squares fit module, a second Tanh activation function, and a second MLP module, all connected in series. The linear feature extraction module uses the PatchPartition module to perform dimensionality blending and reshaping and moving window segmentation on the input linear data. Then, it uses an encoding-decoding structure composed of two MLP modules. Between the two MLP modules, a Least-squares fit module performs multi-step fitting using the least squares method, as expressed by the following formula:

[0044]

[0045] The nonlinear feature extraction module comprises a first ReLU activation function, a third MLP module, a Linear module, a second ReLU module, and a fourth MLP module, which are sequentially connected. The module uses an encoding-decoding structure composed of two MLP modules to process the input nonlinear data. Feature extraction is performed between the two MLP modules using a fully connected layer via a Linear module, as expressed by the following formula:

[0046]

[0047] The MCIE-T branch module is a commonly used single Transformer structure, including a first multi-head attention module, a second multi-head attention module, a masked multi-head attention module, a first residual connection and regularization module, a second residual connection and regularization module, a third residual connection and regularization module, a fifth MLP module, and a sixth MLP module.

[0048] The feature fusion module fuses input feature data at different scales to obtain a multi-scale fused feature map. The feature fusion module includes a first Linear module, a second Linear module, a third Linear module, a first downsampling module, and a second downsampling module. If the feature fusion module uses the output of the short-term prediction branch module as input, the feature fusion process is as follows: the output of the MCIE-T branch module is processed by the first Linear module to obtain the data. The output of the MCIE-L branch module is processed by the second Linear module and then compared with the downsampled data. Add them together to get the data. The output of the MCIE-C branch module is processed by the third Linear module and then compared with the downsampled data. The results are added together and used as the output of the feature fusion module; where F.interpolate(·) is the downsampling operation;

[0049] If the feature fusion module uses the output of the intermediate prediction branch module as input, the feature fusion process is as follows: the output of the MCIE-T branch module is processed by the first Linear module to obtain data. The output of the first MCIE-L branch module is processed by the second Linear module, and then compared with the downsampled data. Add them together to get the data. The output of the second MCIE-L branch module is processed by the third Linear module and then compared with the downsampled data. The results are added together and used as the output of the feature fusion module; where F.interpolate(·) is the downsampling operation;

[0050] If the feature fusion module uses the output of the long-term prediction branch module as input, the feature fusion process is as follows: the output of the first MCIE-T branch module is processed by the first Linear module to obtain data. The output of the second MCIE-T branch module is processed by the second Linear module and then compared with the downsampled data. Add them together to get the data. The output of the third MCIE-T branch module is processed by the third Linear module and then compared with the downsampled data. The results are added together and used as the output of the feature fusion module; where F.interpolate(·) is the downsampling operation;

[0051] The second RevIN module takes the output of the feature fusion module as input, denormalizes the input data, and uses the result as the output of the model.

[0052] The inverse normalization is represented by the following formula:

[0053]

[0054] in, This represents the model's final predicted value; For wind speed data; γ k and β k Let γ be the affine parameter vector to be learned. k ,β k ∈R K .

[0055] This invention discloses a multi-scale cascaded enhancement end-to-end prediction method for wind speed. It adopts a deep learning-based convolutional neural network structure design to achieve accurate prediction of multi-scale wind speed. It has high accuracy, low false detection rate, fast detection speed, and good real-time performance, effectively improving prediction efficiency and showing broad application prospects. Attached Figure Description

[0056] Figure 1 This is a schematic flowchart of the method of the present invention;

[0057] Figure 2 (a-d) show the comparison results of the method of the present invention with other methods under different prediction lengths in the embodiments, wherein the prediction sequence length of Figure a is 96; the prediction sequence length of Figure b is 192; the prediction sequence length of Figure c is 336; and the prediction sequence length of Figure d is 720. Detailed Implementation

[0058] This invention provides a multi-scale cascaded enhancement end-to-end prediction method for wind speed, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps:

[0059] S1. Obtain wind speed data;

[0060] The wind speed data is acquired through wind speed sensors at ground monitoring points and includes temperature parameters, air pressure parameters, and wind speed parameters.

[0061] S2. Preprocess the wind speed data obtained in step S1 to obtain a wind speed dataset, specifically:

[0062] First, the wind speed data obtained in step S1 is preprocessed. The wind speed data detected by the N stations after preprocessing is constructed into a 1*L*W format, where L is the long-term wind speed data recording speed and W is the number of multi-source parameters obtained. Annotations are added to the station where each wind speed data was acquired based on the latitude and longitude relationship of the station. The missing data is filled with the value "0". A prediction judgment parameter λ is added to the wind speed data. The range of the short-term prediction judgment parameter is 0 < λ ≤ 96; the range of the medium-term prediction judgment parameter is 96 < λ ≤ 192; and the range of the long-term prediction judgment parameter is λ > 192.

[0063] The preprocessing includes data missing completion, data anomaly handling, and data standardization. Regarding data missing, forward imputation is used to complete the missing data based on the discontinuous nature of the missing data. Regarding data anomalies, a probabilistic autoencoder-based method is selected to process the data. Regarding data standardization, Z-score standardization is used to process the data, taking into account the inconsistency between wind speed and wind vector dimensions and the requirements of the training model.

[0064] S3. Construct an initial wind speed prediction model based on a multi-scale prediction strategy;

[0065] The initial wind speed prediction model includes a first RevIN module, a judgment module, a short-term prediction branch module, a medium-term prediction branch module, a long-term prediction branch module, a feature fusion module, and a second RevIN module.

[0066] The multi-scale strategy is as follows: first, scale segmentation is performed to generate multi-scale data; then, short-term prediction branch modules, medium-term prediction branch modules, and long-term prediction branch modules with different feature extraction capabilities are used for different scale data; finally, the multi-scale prediction features are fused through a feature fusion module. The model built based on the multi-scale strategy has the ability to flexibly model the dependencies between different prediction lengths across scales.

[0067] The data input to the model is first normalized by the first RevIN module before entering the judgment module; the judgment module makes the following judgments based on the predicted judgment parameters:

[0068] If the prediction judgment parameter 0 < λ ≤ 96, the output of the judgment module enters the short-term prediction branch module; after being processed by the short-term prediction module, it is then fused through the feature fusion module, and the output result is inversely normalized through the second RevIN module to obtain the model output result.

[0069] If the prediction judgment parameter 96 < λ ≤ 192, the output of the judgment module enters the intermediate prediction branch module; after being processed by the intermediate prediction module, the feature fusion module performs feature fusion, and the output result is inversely normalized by the second RevIN module to obtain the model output result.

[0070] If the prediction judgment parameter λ > 192, the output of the judgment module enters the long-term prediction branch module; after being processed by the long-term prediction module, it is then fused through the feature fusion module, and the output result is inversely normalized through the second RevIN module to obtain the model output result.

[0071] The first RevIN module normalizes the input data by calculating the mean and standard deviation, expressed using the following formula:

[0072]

[0073] in, L represents the average value of the input data. p The length of the input data; For each data value; j represents data from different dimensions at the same time, such as dimensions including temperature and air pressure; The deviation of the input data; after obtaining the mean and standard deviation of the data, the input data is normalized using the following formula:

[0074]

[0075] in, The input data is normalized; ε is the disturbance error; the normalized input data is input into the judgment module.

[0076] The short-term prediction branch module includes the MCIE-C branch module, the MCIE-L branch module, and the MCIE-T branch module;

[0077] The intermediate-term prediction branch module includes an MCIE-T branch module, a first MCIE-L branch module, and a second MCIE-L branch module; the first MCIE-L branch module and the second MCIE-L branch module have the same structure.

[0078] The long-term prediction branch module includes a first MCIE-T branch module, a second MCIE-T branch module, and a third MCIE-T branch module; the first to third MCIE-T branch modules have the same structure.

[0079] The MCIE-C branch module includes a first convolutional coding module, a first 1×1 convolutional layer module, a 3×3 depthwise separable convolutional layer module, a concat module, a second 1×1 convolutional layer module, and a second convolutional coding module, which are connected in sequence.

[0080] The MCIE-C branch module performs dimensionality transformation on the input data through the first convolutional coding module, obtains local features through the first 1×1 convolutional layer module and the 3×3 depthwise separable convolutional layer module, and finally exchanges dimensionality information through the second 1×1 convolutional layer module to integrate fine-grained features; the output of the second 1×1 convolutional layer module is processed by the second convolutional coding module and used as the output of the MCIE-C branch module.

[0081] The first 1×1 convolutional layer module has the same structure as the second 1×1 convolutional layer module, including a 1×1 convolutional layer, a ReLU activation function, and a post-activation BatchNorm connected in series.

[0082] The first convolutional coding module performs dimensionality transformation on the input data. It uses a sliding window approach, moving along the time dimension of the input data according to the window length and stride, and extracting subsequence patches as the output. This is expressed by the following formula:

[0083]

[0084] Where P is the window length and S is the step size;

[0085] A 3×3 depthwise separable convolutional layer module captures the global receptive field and local location features in the input data. The entire process inputs N / 2 patches, and each patch undergoes a separate convolution operation with a kernel to generate N feature maps. The size of the convolution kernel is the same as the patch stride in the first convolutional encoding module, as expressed by the following formula:

[0086]

[0087] Where l is the length of the input data;

[0088] The concat module takes the output of the first 1×1 convolutional layer module and the output of the 3×3 depth-separable convolutional layer module as inputs, and concatenates them to create an input with N channels;

[0089] The MCIE-L branch module separates linear and nonlinear features and extracts them separately. The MCIE-L branch module includes a Fourier module, a linear feature extraction module, and a nonlinear feature extraction module. The input data is first processed by the Fourier module to decompose it into linear and nonlinear data. The linear feature extraction module uses the linear data output from the Fourier module as input; the nonlinear feature extraction module uses the nonlinear data output from the Fourier module as input; the output of the MCIE-L branch module is the direct sum of the outputs from the linear and nonlinear feature extraction modules.

[0090] The Fourier module clarifies the sequence components of the input data, calculates and sorts the average amplitude of different data in the frequency domain using the Fast Fourier Transform technique for each past window, and decomposes the input data into linear and nonlinear data, as shown in the following formula:

[0091] X L =FTT -1 (X in )

[0092] X NonL =X in -FTT -1 (X in )

[0093] Among them, X L Linear data; X NonL For non-linear data; X in For the input data of the Fourier module; FTT -1 Inverse Fourier transform;

[0094] The linear feature extraction module comprises a PatchPartition module, a first Tanh activation function, a first MLP module, a Least-squares fit module, a second Tanh activation function, and a second MLP module, all connected in series. The linear feature extraction module uses the PatchPartition module to perform dimensionality blending and reshaping and moving window segmentation on the input linear data. Then, it uses an encoding-decoding structure composed of two MLP modules. Between the two MLP modules, a Least-squares fit module performs multi-step fitting using the least squares method, as expressed by the following formula:

[0095]

[0096] The nonlinear feature extraction module comprises a first ReLU activation function, a third MLP module, a Linear module, a second ReLU module, and a fourth MLP module, which are sequentially connected. The module uses an encoding-decoding structure composed of two MLP modules to process the input nonlinear data. Feature extraction is performed between the two MLP modules using a fully connected layer via a Linear module, as expressed by the following formula:

[0097]

[0098] The MCIE-T branch module is a commonly used single Transformer structure, including a first multi-head attention module, a second multi-head attention module, a masked multi-head attention module, a first residual connection and regularization module, a second residual connection and regularization module, a third residual connection and regularization module, a fifth MLP module, and a sixth MLP module.

[0099] The feature fusion module fuses input feature data at different scales to obtain a multi-scale fused feature map. The feature fusion module includes a first Linear module, a second Linear module, a third Linear module, a first downsampling module, and a second downsampling module. If the feature fusion module uses the output of the short-term prediction branch module as input, the feature fusion process is as follows: the output of the MCIE-T branch module is processed by the first Linear module to obtain the data. The output of the MCIE-L branch module is processed by the second Linear module and then compared with the downsampled data. Add them together to get the data. The output of the MCIE-C branch module is processed by the third Linear module and then compared with the downsampled data. The results are added together and used as the output of the feature fusion module; where F.interpolate(·) is the downsampling operation;

[0100] If the feature fusion module uses the output of the intermediate prediction branch module as input, the feature fusion process is as follows: the output of the MCIE-T branch module is processed by the first Linear module to obtain data. The output of the first MCIE-L branch module is processed by the second Linear module, and then compared with the downsampled data. Add them together to get the data. The output of the second MCIE-L branch module is processed by the third Linear module and then compared with the downsampled data. The results are added together and used as the output of the feature fusion module; where F.interpolate(·) is the downsampling operation;

[0101] If the feature fusion module uses the output of the long-term prediction branch module as input, the feature fusion process is as follows: the output of the first MCIE-T branch module is processed by the first Linear module to obtain data. The output of the second MCIE-T branch module is processed by the second Linear module and then compared with the downsampled data. Add them together to get the data. The output of the third MCIE-T branch module is processed by the third Linear module and then compared with the downsampled data. The results are added together and used as the output of the feature fusion module; where F.interpolate(·) is the downsampling operation;

[0102] The second RevIN module takes the output of the feature fusion module as input, denormalizes the input data, and uses the result as the output of the model.

[0103]

[0104] in, This represents the model's final predicted value; For wind speed data; γ k and β k Let γ be the affine parameter vector to be learned. k ,β k ∈R K .

[0105] S4. Using the wind speed dataset obtained in step S2, train the initial wind speed prediction model obtained in step S3 to obtain the wind speed prediction model.

[0106] S5. Use the wind speed prediction model obtained in step S4 to make actual wind speed predictions.

[0107] The method of the present invention will be further described below with reference to an embodiment:

[0108] Using the method of this invention and commonly available state-of-the-art (SOTA) models, predictions of different prediction lengths were performed on publicly available wind speed datasets. The results are as follows: Figure 2 As shown in the image, the prediction results obtained by the method of this invention best match the actual wind speed data curve.

Claims

1. A multi-scale cascaded enhancement end-to-end prediction method for wind speed, characterized in that, Includes the following steps: S1. Obtain wind speed data; S2. Preprocess the wind speed data obtained in step S1 to obtain a wind speed dataset; S3. Construct an initial wind speed prediction model based on a multi-scale prediction strategy; S4. Using the wind speed dataset obtained in step S2, train the initial wind speed prediction model obtained in step S3 to obtain the wind speed prediction model. S5. Use the wind speed prediction model obtained in step S4 to make actual wind speed predictions; The initial wind speed prediction model in step S3 includes a first RevIN module, a judgment module, a short-term prediction branch module, a medium-term prediction branch module, a long-term prediction branch module, a feature fusion module, and a second RevIN module. The multi-scale strategy is as follows: first, scale segmentation is performed to generate multi-scale data; then, short-term prediction branch modules, medium-term prediction branch modules, and long-term prediction branch modules with different feature extraction capabilities are used for different scale data; finally, the multi-scale prediction features are fused through a feature fusion module. The model built based on the multi-scale strategy has the ability to flexibly model the dependencies between different prediction lengths across scales. The short-term prediction branch modules include the MCIE-C branch module, the MCIE-L branch module, and the MCIE-T branch module; The intermediate-term prediction branch module includes an MCIE-T branch module, a first MCIE-L branch module, and a second MCIE-L branch module; the first MCIE-L branch module and the second MCIE-L branch module have the same structure. The long-term prediction branch module includes a first MCIE-T branch module, a second MCIE-T branch module, and a third MCIE-T branch module; the first to third MCIE-T branch modules have the same structure. The MCIE-C branch module includes a first convolutional coding module and a first... Convolutional layer module Depthwise separable convolutional layer module, concat module, second Convolutional layer module, second convolutional encoding module; The MCIE-C branch module performs dimensionality transformation on the input data through the first convolutional coding module. Convolutional layer modules and Depthwise separable convolutional layer modules acquire local features, and finally, through a second... The convolutional layer module exchanges dimensional information and integrates fine-grained features; second The output of the convolutional layer module is processed by the second convolutional coding module and then used as the output of the MCIE-C branch module. The MCIE-L branching module separates linear and nonlinear features and extracts them separately. The MCIE-L branching module includes a Fourier module, a linear feature extraction module, and a nonlinear feature extraction module. The input data is first processed by the Fourier module to decompose it into linear and nonlinear data. The linear feature extraction module uses the linear data output by the Fourier module as input; the nonlinear feature extraction module uses the nonlinear data output by the Fourier module as input; the output of the MCIE-L branch module is the direct sum of the outputs of the linear feature extraction module and the nonlinear feature extraction module. The MCIE-T branch module is a commonly used single Transformer structure, including a first multi-head attention module, a second multi-head attention module, a masked multi-head attention module, a first residual connection and regularization module, a second residual connection and regularization module, a third residual connection and regularization module, a fifth MLP module, and a sixth MLP module. The feature fusion module fuses input feature data at different scales to obtain a multi-scale fused feature map. The feature fusion module includes a first Linear module, a second Linear module, a third Linear module, a first downsampling module, and a second downsampling module. If the feature fusion module uses the output of the short-term prediction branch module as input, the feature fusion process is as follows: the output of the MCIE-T branch module is processed by the first Linear module to obtain the data. The output of the MCIE-L branch module is processed by the second Linear module and then compared with the downsampled data. Add them together to get the data. The output of the MCIE-C branch module is processed through the third Linear module, and then compared with the downsampled data. The components are added together, and the result is used as the output of the feature fusion module; where, This is for downsampling operations; If the feature fusion module uses the output of the intermediate prediction branch module as input, the feature fusion process is as follows: the output of the MCIE-T branch module is processed by the first Linear module to obtain data. The output of the first MCIE-L branch module is processed by the second Linear module and then combined with the downsampled data. Add them together to get the data. The output of the second MCIE-L branch module is processed by the third Linear module and then compared with the downsampled data. The components are added together, and the result is used as the output of the feature fusion module; where, This is for downsampling operations; If the feature fusion module uses the output of the long-term prediction branch module as input, the feature fusion process is as follows: the output of the first MCIE-T branch module is processed by the first Linear module to obtain data. The output of the second MCIE-T branch module is processed by the second Linear module and then compared with the downsampled data. Add them together to get the data. The output of the third MCIE-T branch module is processed by the third Linear module and then compared with the downsampled data. The two components are added together, and the result is used as the output of the feature fusion module. This is for downsampling operations.

2. The wind speed multi-scale cascaded enhancement end-to-end prediction method according to claim 1, characterized in that, The wind speed data mentioned in step S1 is obtained through wind speed sensors at ground monitoring points, including temperature parameters, air pressure parameters, and wind speed parameters.

3. The wind speed multi-scale cascaded enhancement end-to-end prediction method according to claim 1, characterized in that, Step S2 specifically involves: first, preprocessing the wind speed data obtained in step S1, and then processing the preprocessed data... Wind speed data detected by each station is used to construct... The format, in which, To record long speeds for wind speed data, To ensure the quantity of multi-source parameters obtained, annotations were added to the stations where each wind speed data point was acquired, based on their latitude and longitude relationships. Areas without data were filled with the value "0". Prediction parameters were then added to the wind speed data. The range of parameters for short-term forecasting is: The range of parameters for medium-term forecasting is as follows: The range of parameters for long-term forecasting is as follows: .

4. The wind speed multi-scale cascaded enhancement end-to-end prediction method according to claim 1, characterized in that, The data input to the model is first normalized by the first RevIN module before entering the judgment module; the judgment module makes the following judgments based on the predicted judgment parameters: If the prediction judgment parameters If the output of the judgment module enters the short-term prediction branch module, after processing by the short-term prediction module, feature fusion is performed by the feature fusion module, and the output result is denormalized by the second RevIN module to obtain the model output result. If the prediction judgment parameters If the result is not found, the output of the judgment module enters the intermediate prediction branch module; after processing by the intermediate prediction module, feature fusion is performed by the feature fusion module, and the output result is inversely normalized by the second RevIN module to obtain the model output result. If the prediction judgment parameters If the output of the judgment module is entered into the long-term prediction branch module, after processing by the long-term prediction module, feature fusion is performed by the feature fusion module, and the output result is denormalized by the second RevIN module to obtain the model output result.

5. The wind speed multi-scale cascaded enhancement end-to-end prediction method according to claim 4, characterized in that, The first RevIN module normalizes the input data by calculating the mean and standard deviation, expressed using the following formula: in, This represents the average value of the input data. The length of the input data; For each data value; This refers to data from different dimensions at the same time, such as dimensions including temperature and air pressure. The deviation of the input data; after obtaining the mean and standard deviation of the data, the input data is normalized using the following formula: in, The input data is normalized; To account for disturbance error, normalized input data is input into the judgment module; The second RevIN module has the same structure as the first RevIN module; the second RevIN module takes the output of the feature fusion module as input, performs inverse normalization on the input data, and uses the result as the output of the model. The inverse normalization is represented by the following formula: in, This represents the model's final predicted value; This is wind speed data; and Let be the affine parameter vector to be learned and .

6. The wind speed multi-scale cascaded enhancement end-to-end prediction method according to claim 1, characterized in that, First Convolutional layer module and second The convolutional layer modules have the same structure, including sequentially connected... Convolutional layers, ReLU activation function, and post-activation BatchNorm; The first convolutional coding module performs dimensionality transformation on the input data. It uses a sliding window approach, moving along the time dimension of the input data according to the window length and stride, and extracting subsequence patches as the output. This is expressed by the following formula: in, The length of the window; Step size; Depthwise separable convolutional layer modules capture the global receptive field and local location features in the input data, and the entire process input Each patch undergoes a separate convolution operation with a kernel, generating... Each feature map has a convolutional kernel size that is the same as the Patch stride in the first convolutional coding module, expressed using the following formula: in, The length of the input data; The concat module uses the first... The output of the convolutional layer module and The outputs of depthwise separable convolutional layer modules are used as inputs, and they are concatenated to build a system with... Input to each channel.

7. The wind speed multi-scale cascaded enhancement end-to-end prediction method according to claim 6, characterized in that, The Fourier module clarifies the sequence components of the input data, calculates and sorts the average amplitude of different data in the frequency domain using the Fast Fourier Transform technique for each past window, and decomposes the input data into linear and nonlinear data, as shown in the following formula: in, Linear data; It is non-linear data; This is the input data for the Fourier module; Inverse Fourier transform; The linear feature extraction module comprises a Patch Partition module, a first Tanh activation function, a first MLP module, a Least-squares fit module, a second Tanh activation function, and a second MLP module, all connected in series. The linear feature extraction module uses the Patch Partition module to perform dimensionality mixing and reshaping and moving window segmentation on the input linear data. Then, it uses an encoding / decoding structure composed of two MLP modules. Between the two MLP modules, a Least-squares fit module performs multi-step fitting using the least squares method, as expressed by the following formula: The nonlinear feature extraction module comprises a first ReLU activation function, a third MLP module, a Linear module, a second ReLU module, and a fourth MLP module, which are sequentially connected. The module uses an encoding-decoding structure composed of two MLP modules to process the input nonlinear data. Feature extraction is performed between the two MLP modules using a fully connected layer via a Linear module, as expressed by the following formula: 。

Images