Space-time wind power prediction method based on adaptive wavelet decomposition and improved Transform

Abstract

The invention belongs to the technical field of wind power prediction, and particularly relates to a space-time wind power prediction method based on adaptive wavelet decomposition and improved Transform. The method comprises the following steps: firstly, acquiring a wind power sequence of a target wind power plant in a historical period and feature sequences influencing the wind power, performing preprocessing, calculating correlation between each feature sequence and the wind power sequence, reserving the feature sequence with high correlation with the wind power, and performing normalization processing on the wind power sequence and the reserved feature sequence; then, extracting frequency domain features of the wind power sequence and each feature sequence; forming a feature by the wind power sequence and the frequency domain features of each feature sequence, and performing periodic embedding and spatial embedding on the feature to obtain a model input feature; and finally, constructing a wind power prediction model based on a Transform network, including a U-Net network, inverse Fourier transform and a prediction layer. And inputting the input features into the U-Net network for feature extraction, performing inverse Fourier transform on the output features of the U-Net network, and then obtaining a wind power prediction value through a prediction layer. According to the method, the problems of high noise, high volatility, non-stability and the like of wind power data are solved, and the prediction precision is improved.

Description

Technical Field

[0001] The invention relates to the technical field of wind power prediction, and in particular to a spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer. Background Art

[0002] The power generation capacity of wind turbines is affected by many factors such as wind speed, wind direction and time, and shows obvious randomness, volatility and intermittency, which increases the difficulty of wind power management. The accurate prediction of wind power can greatly reduce this uncertainty and plays an important role in achieving high wind power penetration rate grid connection, maintaining the balance of electricity supply and demand, and improving the operating efficiency and economy of the power system.

[0003] Methods based on deep learning have shown great advantages in wind power forecasting. Research has proved the feasibility of Transformer networks in wind power forecasting tasks. It achieves effective sequence feature learning through a highly parallelized self-attention mechanism, with faster training speed and stronger long-term dependency modeling capabilities. However, the self-attention used in traditional Transformer networks uses point-by-point dot product calculations, which not only has high computational complexity, but also lacks good perception of local features of time series, and has limitations in long-term dependency modeling. Therefore, the long-term prediction ability and local feature extraction ability of traditional Transformer networks are poor, and it is difficult to achieve ideal results in wind power forecasting tasks. At the same time, the Transformer-based wind power forecasting method faces the challenge of the inherent non-stationarity of wind power. Some studies pre-process the wind power time series through stabilization methods such as normalization and differentiation to eliminate the non-stationarity of the wind power time series, but this leads to over-stabilization problems, making it difficult for the model to capture important trends, seasonality and other time information in the time series.

[0004] In view of the above problems, the present invention proposes a spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer. Summary of the invention

[0005] In view of the deficiencies in the prior art, the technical problem to be solved by the present invention is to provide a spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer.

[0006] The present invention solves the technical problem by adopting the following technical solution:

[0007] A spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer includes the following steps:

[0008] Step 1: Obtain the wind power sequence of the target wind farm in the historical period and the characteristic sequence that affects wind power and perform preprocessing, calculate the correlation between each characteristic sequence and the wind power sequence, retain the characteristic sequence with high correlation with wind power, and normalize the wind power sequence and the retained characteristic sequence;

[0009] Step 2: Extract the frequency domain features of wind power sequence and each characteristic sequence;

[0010] The wind power sequence is decomposed step by step using the adaptive multi-level wavelet decomposition algorithm to obtain multiple high-frequency components and one low-frequency component. The decomposition process is expressed as:

[0011]

[0012] A0=X origin (3)

[0013]

[0014] Where D k , A k is the high-frequency component and low-frequency component obtained by the k-th level decomposition, A k-1 is the low-frequency component obtained by the k-1th level decomposition, A0 is the initial low-frequency component, X origin is the wind power sequence, σ(·) represents the sigmoid activation function, is the learnable weight matrix and bias of the k-th level low-pass filter, is the learnable weight matrix and bias of the k-th level high-pass filter, ε represents the learnable parameters, l1,l2,...,l c are the coefficients of the low-pass filter, h1,h2,...,h c is the coefficient of the high-pass filter, c is the number of coefficients;

[0015] Each high-frequency component and low-frequency component is passed through a frequency domain normalization module to obtain a frequency domain normalized high-frequency component and a low-frequency component; all the frequency domain normalized high-frequency components and low-frequency components are cascaded in the feature dimension to obtain the frequency domain characteristics of the wind power sequence;

[0016] Similarly, the frequency domain characteristics of each feature sequence are obtained;

[0017] Step 3: The frequency domain features of the wind power sequence and each feature sequence form feature X, and feature X is periodically embedded and spatially embedded to obtain feature X in ;

[0018] Step 4: Build a wind power prediction model based on the Transformer network, including the U-Net network, inverse Fourier transform and prediction layer; transform the feature X in The output features of the U-Net network are input into the U-Net network for feature extraction. After the output features of the U-Net network are inverse Fourier transformed, they are passed through the prediction layer to obtain the wind power prediction value.

[0019] The U-Net network includes multiple spatiotemporal encoder layers and spatiotemporal decoder layers, each spatiotemporal encoder layer is jump-connected with the corresponding spatiotemporal decoder layer; the spatiotemporal encoder layer and the spatiotemporal decoder layer both include spatiotemporal attention blocks, layer normalization and feedforward neural networks, the input features are residually connected with themselves after passing through the spatiotemporal attention blocks, the features obtained by the residual connection are used as input features of the feedforward neural network after layer normalization, the input features of the feedforward neural network are residually connected with the output features and then normalized by layers to obtain the output features of the spatiotemporal encoder layer or the spatiotemporal decoder layer;

[0020] Step 5: Train the wind power prediction model, update the parameters of the wind power prediction model, high-pass filter and low-pass filter through back propagation, obtain the trained wind power prediction model, and use the trained wind power prediction model for wind power prediction.

[0021] Furthermore, during the model training process, the training loss is calculated by the stationary correction loss function; the stationary correction loss function L is expressed as:

[0022] L=L1+λL2 (12)

[0023]

[0024] Where L1 is the Huber loss, L2 is the data distribution constraint loss, λ and δ are hyperparameters, and Y, are the true value and predicted value of wind power series, R(Y), is the true value Y and predicted value of the wind power series The autocorrelation matrix of , ||·|| F is the F norm, Conv(·) is the covariance calculation function; Y i , Y j They represent the wind power sequence starting at time i and time j, i,j∈[t+1,t+ΔT], t is the current time, ΔT is the time step, μ i , μ j Respectively represent Y i and Y j The variance of .

[0025] Furthermore, the spatiotemporal attention block includes a multi-head self-attention unit, a causal convolutional self-attention unit and a gated fusion unit; the multi-head self-attention unit uses multi-head self-attention to extract spatial features, the causal convolutional self-attention unit uses causal convolutional self-attention to extract temporal features, and the gated fusion unit fuses spatial features and temporal features to obtain spatiotemporal fusion features; the fusion process of the gated fusion unit is expressed as:

[0026] X ST,out =z⊙X S +(1-z)⊙X T (9)

[0027] z=σ(X S W z,1 +X F W z,2 +b z ) (10)

[0028] In the formula, X ST,out is the spatiotemporal fusion feature, X s , X T They are spatial features and temporal features respectively, ⊙ represents the element-by-element dot product, W z,1 and W z,2 is a learnable weight parameter, b z is bias.

[0029] Furthermore, define the intraday embedding matrix, daily embedding matrix, monthly embedding matrix and spatial embedding matrix, extract the intraday embedding features, daily embedding features and monthly embedding features from the intraday embedding matrix, daily embedding matrix and monthly embedding matrix respectively; concatenate the feature X, the intraday embedding features, the daily embedding features, the monthly embedding features and the spatial embedding features in the feature dimension to obtain the feature X in .

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. In view of the problems of high noise, high volatility and non-stationarity faced by the existing spatiotemporal wind power forecast in processing the original time series data, the different characteristic modes of the sequence are mixed in the time domain, and it is difficult to effectively extract multiple characteristic modes at different resolutions by relying solely on time domain analysis. In order to solve these problems, the present invention improves the wavelet decomposition algorithm, adaptively adjusts the wavelet decomposition coefficients during the training process, and optimizes the wavelet decomposition process. This adaptive characteristic enables the model to maintain a high prediction accuracy under different environments and conditions. Through wavelet decomposition, each sequence is decomposed into low-frequency components and high-frequency components, which can effectively extract multi-level time-frequency features and fully explore the characteristics of the sequence, thereby improving noise resistance and making the model more robust when processing high-noise data.

[0032] 2. When frequency decomposition is applied to time series and input into the Transformer model for feature extraction, frequency deviation problem will occur, causing the model to mainly learn low-frequency features in the sequence and ignore high-frequency features, so that the model pays uneven attention to frequency features with higher energy and cannot accurately capture important high-frequency features. Therefore, the present invention proposes a frequency domain normalization module, which normalizes the low-frequency components and high-frequency components in the frequency domain and then inputs them into the model for feature extraction, thereby reducing the model's bias towards low-frequency features and ensuring that important high-frequency features can also be effectively captured, so that the model can learn features equally in different frequency bands to alleviate the frequency deviation problem, which is beneficial to improving the accuracy of wind power prediction.

[0033] 3. In view of the problem that the self-attention mechanism has limited long-term modeling capabilities when extracting temporal features, the causal attention mechanism is introduced into the spatiotemporal encoder layer and the spatiotemporal decoder layer of the present invention to extract temporal features, thereby better extracting the temporal features of long time series, avoiding the computational burden caused by the increase in sequence length in the self-attention mechanism, and improving the local feature extraction capability of the model. The self-attention mechanism is used to extract spatial features, which enhances the ability to capture global spatial correlations, thereby more comprehensively understanding the spatiotemporal characteristics of wind power series. In view of the problem that traditional Transformer cannot handle the spatial correlation of wind power time series data and lacks long-term modeling capabilities, the prediction model of the present invention enhances the model's ability to capture long-term dependencies by aggregating features of different time scales, can extract and integrate information at different levels, extract spatiotemporal features through the spatiotemporal encoder-decoder layer, and capture multi-scale long-term correlations and global spatial correlations at the same time.

[0034] 4. The non-stationarity of wind power time series hinders the propagation of deep stable features, destroys the data distribution and causes the problem of over-stationarity in previous stabilization methods. The stationarity correction loss function of the present invention can effectively deal with the data distribution offset problem caused by non-stationarity, avoids the over-stationarity phenomenon that may be caused by traditional stabilization methods, maintains time dependence, and thus improves the prediction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall flow chart of the present invention;

[0036] Figure 2 It is a structural diagram of the wind power prediction model of the present invention;

[0037] Figure 3 It is a structural diagram of the U-Net network of the present invention;

[0038] Figure 4 is a structural diagram of the spatiotemporal encoder layer of the present invention;

[0039] Figure 5 It is the mean absolute error (MAE) curve of different models;

[0040] Figure 6 Root mean square error (RMSE) curves of different models;

[0041] Figure 7 It is a comparison chart of the wind power prediction value and the actual value of the present invention. DETAILED DESCRIPTION

[0042] Specific embodiments are given below in conjunction with the accompanying drawings. The specific embodiments are only used to introduce the technical solutions of the present invention in detail and are not intended to limit the protection scope of the present application.

[0043] The present invention provides a spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer (hereinafter referred to as method, see Figures 1 to 7 ), including the following steps:

[0044] Step 1: Obtain the wind power sequence of the target wind farm in the historical period and the characteristic sequence that affects wind power, and pre-process each sequence; calculate the correlation between each characteristic sequence and the wind power sequence, and retain the characteristic sequence with high correlation with wind power; normalize the wind power sequence and the retained characteristic sequence;

[0045] Preprocessing includes outlier processing and missing value filling. The Pearson correlation coefficient method is used to evaluate the correlation between each feature sequence and the wind power sequence. The feature sequences with a correlation coefficient greater than 0.2 are retained, and the features with low correlation with wind power are eliminated, thereby reducing the complexity of model training and improving the training speed and prediction accuracy. The wind power sequence and the retained feature sequences are normalized to eliminate the potential impact of dimensional differences between different parameters on the prediction results.

[0046] Step 2: Extract the frequency domain features of wind power sequence and each characteristic sequence;

[0047] The wind power sequence and each retained feature sequence are decomposed by using an adaptive multi-level wavelet decomposition algorithm, and each sequence is decomposed into multiple high-frequency components and one low-frequency component. Each high-frequency component and low-frequency component are passed through a frequency domain normalization module to obtain frequency domain normalized high-frequency components and low-frequency components; all frequency domain normalized high-frequency components and low-frequency components of the same sequence are cascaded in the feature dimension to obtain the frequency domain features of the sequence;

[0048] Taking the wind power sequence as an example, it is assumed that the wind power sequence is decomposed into K high-frequency components D1, D2, …D by the adaptive multi-level wavelet decomposition algorithm. K and a low-frequency component A KThe low-frequency component corresponds to the trend information of wind power, reflecting the overall and long-term trend changes of wind power; the high-frequency component corresponds to the detailed information of wind power, reflecting the rapid fluctuation and local changes of wind power. All high-frequency and low-frequency components are input into the frequency domain normalization module, and are first converted from the time domain to the frequency domain through Fourier transform to obtain the high-frequency component after Fourier transform. and low frequency components High frequency components after Fourier transform and low frequency components After normalization, the high-frequency component normalized in the frequency domain is obtained. and low frequency components The high-frequency component after normalization in the frequency domain and low frequency components Cascading in the feature dimension, the frequency domain characteristics of the wind power sequence are obtained Among them, || represents a cascade operation;

[0049] The process of adaptive multi-level wavelet decomposition is: using db4 wavelet to initialize the high-pass filter and low-pass filter, and using the high-pass filter and low-pass filter to decompose the wind power sequence step by step. The decomposition process is expressed as:

[0050]

[0051]

[0052] A0=X origin (3)

[0053]

[0054] Where D k , A k is the high-frequency component and low-frequency component obtained by the k-th level decomposition, A k-1 is the low-frequency component obtained by the k-1th level decomposition, A0 is the initial low-frequency component, X origin is the wind power sequence, σ(·) represents the sigmoid activation function, is the learnable weight matrix and bias of the k-th level low-pass filter, is the learnable weight matrix and bias of the k-th level high-pass filter; ε represents the learnable parameter, initialized to a value close to zero; l1,l2,...,l c are the coefficients of the low-pass filter, h1,h2,...,h c is the coefficient of the high-pass filter, c is the number of coefficients;

[0055] Step 3: The frequency domain features of the wind power sequence and each retained feature sequence form feature X = [x1, x2, ..., x r ] T ; where x1,x2,...,x r is the frequency domain feature of each sequence, r is the number of sequences; the feature X is input into the embedding layer to realize periodic embedding and spatial embedding, so as to extract the spatiotemporal correlation of each time series and obtain the feature X in ;

[0056] Step 3.1: Obtain intra-day embedding features, daily embedding features, and monthly embedding features;

[0057] Defining a learnable intra-day embedding matrix Daily Embedding Matrix and the monthly embedding matrix To store the time characteristics of information; where N t Indicates the number of samples in a day, N d Indicates the number of days in a year, N m Indicates the month number, d t ,d d ,d m Represent the intraday, daily, and monthly embedding dimensions respectively; through the intraday, daily, and monthly time information corresponding to the wind power time series of the N wind turbines in the wind farm at time t, the intraday embedding features are extracted from the intraday embedding matrix, the daily embedding matrix, and the monthly embedding matrix respectively. Daily Embedding Features and monthly embedding features

[0058] Step 3.2: Define spatial embedding features Among them, d s Represents the spatial embedding dimension;

[0059] Step 3.3: Embed feature X and intraday into feature E t , daily embedding feature E d , monthly embedding feature E m and spatial embedding feature E s Cascade in the feature dimension to get feature X in =X||E t ||E d ||E m ||E s , which are the input features of the model.

[0060] Step 4: Build a wind power prediction model based on the Transformer network, including the U-Net network, inverse Fourier transform, and prediction layer;

[0061] The U-Net network includes an encoder and a decoder, the encoder includes multiple spatiotemporal encoder layers and spatiotemporal decoder layers, each spatiotemporal encoder layer is jump-connected with the corresponding spatiotemporal decoder layer; the output features of the previous spatiotemporal encoder layer are down-sampled to obtain the input features of the next spatiotemporal encoder layer; the output features of the previous spatiotemporal decoder layer are up-sampled to obtain the input features of the next spatiotemporal decoder layer.

[0062] The architecture of the spatiotemporal encoder layer and the spatiotemporal decoder layer is the same, both including spatiotemporal attention blocks, layer normalization and feedforward neural networks; the input features are residually connected with themselves after passing through the spatiotemporal attention block, and the features obtained by the residual connection are input into the feedforward neural network after layer normalization. The input features of the feedforward neural network are residually connected with the output features and then layer normalized to obtain the output features of the spatiotemporal encoder layer or the spatiotemporal decoder layer.

[0063] The spatiotemporal attention block includes a multi-head self-attention unit, a causal convolution self-attention unit and a gated fusion unit. The multi-head self-attention unit is used to extract spatial features, the causal convolution self-attention unit is used to extract temporal features, and the gated fusion unit is used to fuse spatial features and temporal features to obtain spatiotemporal fusion features, that is, the output features of the spatiotemporal attention block; in the multi-head self-attention unit, the input feature X of the spatiotemporal attention block ST,in After multi-head self-attention, in each head, the query vector, key vector and value vector are first obtained by linear projection; among them, the query vector Q of the h-th head h , key vector K h Sum value vector V h Calculated by the following formula:

[0064]

[0065] In the formula, and are all weight matrices, H is the number of attention heads;

[0066] The query vector, key vector and value vector are calculated using the standard dot product, and the attention calculation of the h-th head is expressed as:

[0067]

[0068] In the formula, head h is the attention calculation result of the h-th head, softmax(·) is the softmax activation function, and d k is the dimension of the key vector;

[0069] The multi-head attention calculation results are cascaded and then linearly projected to obtain the output features of the multi-head self-attention unit, that is, the spatial feature X S ;

[0070] X S =(head1||head2||...||head H )W s (8)

[0071] Where W s is the weight matrix;

[0072] In the spatiotemporal attention block, the temporal feature X is extracted using the causal convolution self-attention unit T In the standard dot product self-attention mechanism, the similarity between the query vector and the key vector is calculated based on the point-by-point dot product, which does not fully utilize the local contextual information in the sequence. The causal convolutional attention introduces a causal convolutional layer, which performs convolution operations on the time series through a one-dimensional convolution kernel. The output of each layer is obtained by combining the hidden input of the previous layer and the input information of the previous position, and ensures that the current position does not obtain future information.

[0073] Gated fusion unit for spatial feature X s and time feature X T Fusion is performed to obtain the spatiotemporal fusion feature, that is, the output feature X of the spatiotemporal attention block ST,out ∈R L×d ; The calculation formula is:

[0074] X ST,out =z⊙X S +(1-z)⊙X T (9)

[0075] z=σ(X S W z,1 +X F W z,2 +b z ) (10)

[0076] Where W z,1 ∈R d×d and W z,2 ∈R d×d is a learnable weight parameter, b z ∈R d is the bias and ⊙ represents the element-wise dot product.

[0077] The output features of the U-Net network are inversely transformed to the time domain. The features after inverse Fourier transformation are passed through the prediction layer to obtain the wind power prediction value. The prediction layer consists of two fully connected layers, and the wind power prediction value is calculated by the following formula:

[0078]

[0079] In the formula, is the wind power forecast value for the next ΔT time steps, X O is the feature after inverse Fourier transform, W1∈R L×ΔT , W2∈R d×1 is a learnable weight parameter, b1 and b2 are biases.

[0080] Step 5: Train the wind power prediction model, calculate the training loss through the following stability correction loss function, update the parameters of the wind power prediction model, high-pass filter and low-pass filter through back propagation, obtain the trained wind power prediction model, and use the trained wind power prediction model for wind power prediction;

[0081] The stationarity correction loss function L is:

[0082] L=L1+λL2 (12)

[0083]

[0084] Where L1 is the Huber loss, L2 is the data distribution constraint loss, λ and δ are hyperparameters, and Y, are the true value and predicted value of wind power series, R(Y), is the true value Y and predicted value of the wind power series The autocorrelation matrix of , ||·|| F is the F norm, Conv(·) is the covariance calculation function; Y i , Y j They represent the wind power sequence starting at time i and time j, i,j∈[t+1,t+ΔT], μ i , μ j Respectively represent Y i and Y j The variance of .

[0085] The stationarity correction loss function effectively alleviates the data distribution offset problem caused by non-stationarity. It can not only restore the data distribution and retain the key information in the non-stationarity, but also maintain the time dependence of the wind power series. The data distribution constraint loss is used to regulate the stationarity correlation between the previous and next sequences, realize the correction of data distribution, and improve the robustness of processing non-stationary wind power series.

[0086] In order to verify the effectiveness of the method of the present invention, the model of the present invention is used to predict the wind power in the next 12 steps (2 hours) based on the wind power in the past 24 time steps (4 hours), and is compared with the existing Transformer, Informer, and iTransformer models. The mean absolute error (MAE) and root mean square error (RMSE) are selected to evaluate the model performance. Figure 5 is the mean absolute error curve of each model, Figure 6 is the root mean square error curve of each model, Figure 7 It is a comparison between the wind power prediction value of the present invention and the actual value.

[0087] As can be seen from the figure, the MAE and RMSE of the method of the present invention at each time step are significantly lower than those of other models, and the wind power prediction value is closer to the true value, with good prediction performance. This result shows that the model of the present invention performs well in extracting the complex characteristics of high-volatility and non-stationary wind power sequences. This is because the coefficients of the high-pass filter and the low-pass filter are automatically adjusted by adaptive multi-level wavelet decomposition, effectively extracting multi-level time-frequency features. Frequency deviation is alleviated by frequency domain normalization, and prediction accuracy is improved. The spatiotemporal encoder layer and the spatiotemporal decoder layer enhance the model's ability to capture long-term dependencies, improve the ability to extract local features and process long time series through causal convolution self-attention, handle the data distribution offset caused by non-stationarity through the stationarity correction loss function, restore the data distribution and retain non-stationarity information, while maintaining time dependency, further improving the stability of the model. Therefore, the model of the present invention can effectively capture time domain, spatial domain and frequency domain features, and significantly improve the accuracy and efficiency of wind power prediction.

[0088] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer, characterized in that: The method comprises the following steps: Step 1: Obtain the wind power sequence of the target wind farm in the historical period and the characteristic sequence that affects wind power and perform preprocessing, calculate the correlation between each characteristic sequence and the wind power sequence, retain the characteristic sequence with high correlation with wind power, and normalize the wind power sequence and the retained characteristic sequence; Step 2: Extract the frequency domain features of wind power sequence and each characteristic sequence; The wind power sequence is decomposed step by step using the adaptive multi-level wavelet decomposition algorithm to obtain multiple high-frequency components and one low-frequency component. The decomposition process is expressed as: A0=X origin (3) Where D k , A k is the high-frequency component and low-frequency component obtained by the k-th level decomposition, A k-1 is the low-frequency component obtained by the k-1th level decomposition, A0 is the initial low-frequency component, X origin is the wind power sequence, σ(·) represents the sigmoid activation function, is the learnable weight matrix and bias of the k-th level low-pass filter, is the learnable weight matrix and bias of the k-th level high-pass filter, ε represents the learnable parameters, l1,l2,...,l c are the coefficients of the low-pass filter, h1,h2,...,h c is the coefficient of the high-pass filter, c is the number of coefficients; Each high-frequency component and low-frequency component is passed through a frequency domain normalization module to obtain a frequency domain normalized high-frequency component and a low-frequency component; all the frequency domain normalized high-frequency components and low-frequency components are cascaded in the feature dimension to obtain the frequency domain characteristics of the wind power sequence; Similarly, the frequency domain characteristics of each feature sequence are obtained; Step 3: The frequency domain features of the wind power sequence and each feature sequence form feature X, and feature X is periodically embedded and spatially embedded to obtain feature X in ; Step 4: Build a wind power prediction model based on the Transformer network, including the U-Net network, inverse Fourier transform and prediction layer; transform the feature X in The output features of the U-Net network are input into the U-Net network for feature extraction. After the output features of the U-Net network are inverse Fourier transformed, they are passed through the prediction layer to obtain the wind power prediction value. The U-Net network includes multiple spatiotemporal encoder layers and spatiotemporal decoder layers, each spatiotemporal encoder layer is jump-connected with the corresponding spatiotemporal decoder layer; the spatiotemporal encoder layer and the spatiotemporal decoder layer both include spatiotemporal attention blocks, layer normalization and feedforward neural networks, the input features are residually connected with themselves after passing through the spatiotemporal attention blocks, the features obtained by the residual connection are used as input features of the feedforward neural network after layer normalization, the input features of the feedforward neural network are residually connected with the output features and then normalized by layers to obtain the output features of the spatiotemporal encoder layer or the spatiotemporal decoder layer; Step 5: Train the wind power prediction model, update the parameters of the wind power prediction model, high-pass filter and low-pass filter through back propagation, obtain the trained wind power prediction model, and use the trained wind power prediction model for wind power prediction.

2. The spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer according to claim 1 is characterized in that: During the model training process, the training loss is calculated by the stationary correction loss function; the stationary correction loss function L is expressed as: L=L1+λL2 (12) Where L1 is the Huber loss, L2 is the data distribution constraint loss, λ and δ are hyperparameters, and Y, are the true value and predicted value of wind power series, R(Y), is the true value Y and predicted value of the wind power series The autocorrelation matrix of , ||·|| F is the F norm, Conv(·) is the covariance calculation function; Y i , Y j They represent the wind power sequence starting at time i and time j, i,j∈[t+1,t+ΔT], t is the current time, ΔT is the time step, μ i , μ j Respectively represent Y i and Y j The variance of .

3. The spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer according to claim 1 or 2, characterized in that: The spatiotemporal attention block includes a multi-head self-attention unit, a causal convolution self-attention unit and a gated fusion unit; the multi-head self-attention unit uses multi-head self-attention to extract spatial features, the causal convolution self-attention unit uses causal convolution self-attention to extract temporal features, and the gated fusion unit fuses spatial features and temporal features to obtain spatiotemporal fusion features; the fusion process of the gated fusion unit is expressed as: X ST,out =z⊙X S +(1-z)⊙X T (9) z=σ(X S W z,1 +X F W z,2 +b z ) (10) Where, X ST,out is the spatiotemporal fusion feature, X s , X T They are spatial features and temporal features respectively, ⊙ represents the element-by-element dot product, W z,1 and W z,2 is a learnable weight parameter, b z is bias.

4. The spatiotemporal wind power prediction method based on adaptive wavelet decomposition and improved Transformer according to claim 1 is characterized in that: Define an intraday embedding matrix, a daily embedding matrix, a monthly embedding matrix, and a spatial embedding matrix, and extract intraday embedding features, daily embedding features, and monthly embedding features from the intraday embedding matrix, the daily embedding matrix, and the monthly embedding matrix, respectively; Concatenate feature X, intra-day embedding feature, daily embedding feature, monthly embedding feature, and spatial embedding feature in the feature dimension to obtain feature X in .

Images