Ultra-short-term wind speed prediction method based on matrix spatial-temporal feature decomposition and multi-feature fusion network

Abstract

The invention discloses an ultra-short-term wind speed prediction method based on matrix spatial-temporal feature decomposition and a multi-feature fusion network, and solves the technical problems that the number of effective elements exists in matrix spatial features and local detail information is lost in a dense wind power plant according to a wind speed prediction method constructed based on spatial-temporal features. The method comprises the following steps: acquiring real-time data of each wind turbine in a wind power plant, preprocessing the data, and extracting matrix spatial-temporal characteristics; decomposing the matrix spatial-temporal characteristics by using VMD to obtain a plurality of sub-matrix spatial-temporal characteristics; and inputting sub-matrix spatio-temporal features, and performing prediction through a multi-feature fusion network. According to the method, the overall prediction of the large wind power plant is realized, and the prediction precision is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of wind speed prediction technology, specifically to an ultra-short-term wind speed prediction method based on matrix spatiotemporal feature decomposition and multi-feature fusion network. Background Technology

[0002] Wind energy, as one of the renewable energy sources with the greatest potential for large-scale development, suffers from significant power output fluctuations due to its intermittent and highly random nature, posing a severe challenge to the real-time balance and stable operation of the power system. Against this backdrop, ultra-short-term wind speed forecasting (with a forecast timescale of 15 minutes to 4 hours) has become a core technological support for wind farm power dispatch, equipment maintenance, and grid security. High-precision wind speed forecasting can reduce the impact of wind power grid connection on grid frequency regulation, decrease reserve capacity configuration costs, and simultaneously improve the economic benefits of wind farms.

[0003] Existing wind speed forecasting methods are mainly divided into three categories: physical model-based methods, statistical methods, and data-driven methods. Physical models (such as numerical weather prediction, NWP) rely on atmospheric dynamic equations. Although they are stable in medium- and long-term forecasts, their computational complexity is high and they struggle to capture local micro-meteorological changes, failing to meet the real-time requirements of ultra-short-term forecasts. Statistical methods (such as ARIMA and Kalman filtering) establish linear relationship models using historical data, but they are poorly adapted to the nonlinear and non-stationary characteristics of wind speed. In recent years, data-driven methods, represented by deep learning (such as LSTM, CNN, and Transformer), have gradually become a research hotspot by automatically extracting spatiotemporal features. For example, the paper "Wind speedprediction using a hybrid model combining CNN and GRU" proposes a model that combines convolutional neural networks (CNN) and gated recurrent units (GRU). CNN captures spatial correlations, while GRU learns temporal dependencies. However, in dense wind farms, the geographical distribution of multiple wind turbines is complex, and their spatiotemporal characteristics are easily affected by the wake effect of nearby units and the coupling effect of terrain shading, making it difficult for traditional single-scale feature extraction methods to fully represent local detailed information.

[0004] For modeling spatiotemporal features, existing technologies attempt to map the geographic coordinates of wind turbines into a matrix structure to express spatial relationships. For example, Chinese patent CN 114912577 A discloses a wind speed prediction method combining variational mode decomposition (VMD) and attention mechanisms. This method constructs a wind speed spatiotemporal matrix (SWSM) and decomposes it into subsequences, using SENet to extract spatial features and GRU to capture temporal features. However, this method has two major limitations: First, the non-stationarity of wind speed signals makes Empirical Mode Decomposition (EMD) methods prone to mode aliasing, and the frequency band overlap of subsequences weakens feature discriminative power. Second, the feature fusion stage only uses simple concatenation or weighted superposition, failing to fully consider the complementarity between shallow local features and deep abstract features, leading to the loss of detailed information. Especially in densely populated wind turbine areas, the effective element ratio of matrix spatial features is low, further exacerbating the risk of misjudging sudden wind speed trends. Summary of the Invention

[0005] The purpose of this invention is to provide an ultra-short-term wind speed prediction method based on matrix spatiotemporal feature decomposition and multi-feature fusion network, so as to solve the technical problems of limited effective elements and loss of local detail information in wind speed prediction methods based on spatiotemporal features in dense wind farms.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The ultra-short-term wind speed prediction method based on matrix spatiotemporal eigenvalue decomposition and multi-feature fusion network provided by this invention includes the following steps: Step 1: Obtain real-time data for each wind turbine in the wind farm, and map the data into a two-dimensional matrix based on geographical location information to form matrix spatial features; Step 2: Superimpose the matrix space features at consecutive time points to form a three-dimensional tensor containing time information in the channel direction; Step 3: Decompose the three-dimensional tensor by variational mode decomposition to obtain multiple sub-modes, each sub-mode focusing on different frequency components in the signal; Step 4: Input the decomposed submodalities into the multi-feature fusion network and use the MFFNet network to extract local spatiotemporal features; Step 5: Calculate the global attention matrix through a multi-head attention mechanism, and combine it with residual connections and normalization modules to enhance feature learning capabilities; Step 6: Predict the extracted features using a linear layer and a softmax activation function, and output the normalized prediction power; Step 7: Through inverse normalization, the prediction results are transformed into the actual wind speed prediction value at the target time point.

[0007] Furthermore, the process of constructing the matrix space features in step 1 includes: mapping the real-time data of the wind turbine into a two-dimensional matrix according to its latitude and longitude coordinates, where each element of the matrix represents the state of the wind turbine at a specific location at a certain moment.

[0008] Furthermore, the dimensions of the three-dimensional tensor are Channel × Height × Width, where Channel represents the number of channels, signifying different feature dimensions; Height and Width represent the number of wind turbines in the determined region in the vertical and horizontal directions, respectively, and their expressions are as follows: ; Among them, X n (t) represents the characteristics of the nth wind turbine at time t.

[0009] Furthermore, the variational mode decomposition process in step 3 includes: The input three-dimensional tensor is decomposed into multiple sub-modes by using variational optimization methods. Each sub-mode has an independent frequency center and bandwidth constraint. An optimization algorithm using the alternating direction multiplier method is employed to update the bandwidth and center frequency of each mode in the frequency domain, ensuring the accuracy of the decomposition. The optimization objective of variational mode decomposition is expressed as: ; Where k represents the mode index, X(t) represents the original signal, j represents the imaginary unit, and u j (t) represents the j-th mode function, α represents the equilibrium factor, and δu k δt represents the change in the modal signal, and δt represents the time step.

[0010] Furthermore, in step 5, the local spatiotemporal features extracted by MFFNet are first reshaped into the form of h×w, hidden_dim. This reshaping adopts a row-first approach to expand the matrix into a one-dimensional form, forming a global attention matrix.

[0011] Furthermore, in step 5, the global attention matrix is ​​calculated using a multi-head attention mechanism. The calculation process of the multi-head attention mechanism is as follows: ; Where Q represents the query matrix, K represents the key matrix, V represents the value matrix, T represents the matrix transpose operation, and d k This represents the dimension of the key vector.

[0012] Based on the above technical solution, the embodiments of the present invention can produce at least the following technical effects: (1) The ultra-short-term wind speed prediction method based on matrix spatiotemporal eigenvalue decomposition and multi-feature fusion network provided by this invention proposes a spatiotemporal eigenvalue decomposition method based on variational mode decomposition (VMD). By introducing a variational model and frequency bandwidth constraints, this method can decompose complex signals into multiple modes with different frequency components. This method overcomes the mode aliasing problem in traditional empirical mode decomposition (EMD), enabling each mode to clearly represent specific frequency information in the signal. Compared with existing spatiotemporal eigenvalue decomposition methods, this method not only improves the accuracy of signal decomposition but also provides a reliable data foundation for subsequent feature extraction and wind speed prediction.

[0013] (2) The ultra-short-term wind speed prediction method based on matrix spatiotemporal feature decomposition and multi-feature fusion network provided by this invention addresses the issue that deep features extracted by CNNs represent abstract features of the target object, which ignore detailed information of the target object. The multi-feature fusion network (MFFNet) proposed in this invention fuses shallow, medium, and deep features, compensating for lost local information while focusing on global information, thereby improving the prediction accuracy of the model. Given the different sizes of shallow, medium, and deep features, bilinear interpolation is used to restore the shallow, medium, and deep features to the same resolution as the input image, and they are stitched together in the channel direction. Compared with traditional feature fusion methods, the prediction accuracy is significantly improved. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the ultra-short-term wind speed prediction method of the present invention; Figure 2 This is the matrix spatiotemporal eigenvalue decomposition diagram based on VMD of this invention; Figure 3 This is a schematic diagram of the MFFNet network structure of the present invention. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0017] This embodiment provides an ultra-short-term wind speed prediction method based on matrix spatiotemporal feature decomposition and multi-feature fusion network. The model consists of two parts: a data preprocessing stage based on variational mode decomposition (VMD) matrix spatiotemporal feature decomposition and a prediction stage based on multi-feature fusion network (MFFNet). VMD decomposes the matrix spatiotemporal features into multiple sub-matrix spatiotemporal features, while the prediction part based on the MFFNet multi-feature fusion network learns the intrinsic data features of the spatiotemporal features. These two stages constitute different phases of the wind speed prediction problem.

[0018] VMD predetermines the number of solution subsequences to ensure that different wind speed sequences yield the same number of subsequences, thus reducing wind speed nonlinearity and nonstationarity while constructing spatiotemporal features. This method effectively avoids the mode mixing problem in traditional EMD and can separate clear frequency components when processing signals with significant noise interference.

[0019] In the prediction phase, the power prediction part utilizes MFFNet to extract local spatiotemporal features and global correlations between sites, establishing the relationship between local spatiotemporal features and output wind speed. A schematic diagram of the MFFNet network structure is shown below. Figure 3 As shown, the core structure of MFFNet includes multiple parallel sub-networks, each focusing on learning the intrinsic relationships between spatiotemporal features of different sub-matrices. The MFFNet prediction module is similar to the encoder layer of a Transformer. It first computes a multi-head attention module that calculates the attention matrix between stations in the spatial dimension, and further enhances feature learning capabilities through residual connections and a normalization module. This prediction part can be stacked in multiple layers to capture deeper spatiotemporal features, with the final output dimension being (h×w, hidden_dim), corresponding to the features of h×w stations within the region. A linear layer follows the prediction part, reducing the hidden_dim dimension to 1 and outputting a normalized result through a softmax activation function, resulting in a final dimension of (h×w, 1). The prediction result is multiplied by the equivalent rated power of the corresponding station for inverse normalization to obtain the final wind speed prediction value.

[0020] like Figures 1-3 As shown, the ultra-short-term wind speed prediction method based on matrix spatiotemporal eigenvalue decomposition and multi-feature fusion network specifically includes the following steps: Step 1: Obtain real-time data for each wind turbine in the wind farm, and map the data into a two-dimensional matrix based on geographical location information (latitude and longitude coordinates) to form matrix spatial features; Each element of the matrix represents the state of a wind turbine at a specific location at a certain moment. This process integrates real-time data of wind turbines (such as wind speed, temperature, pressure, etc.) into a matrix, enabling the model to make full use of spatial information and capture the geographical and functional relationships between wind turbines. Step 2: Superimpose the matrix space features at consecutive time points to form a three-dimensional tensor containing time information in the channel direction; The three-dimensional tensor has dimensions of Channel × Height × Width, where Channel represents the number of channels, signifying different feature dimensions; Height represents the vertical dimension; and Width represents the horizontal dimension, corresponding to the number of wind turbines within the region to be determined. Its expression is as follows: ; Where Xn(t) represents the characteristics of the nth wind turbine at time t.

[0021] Step 3: Decompose the three-dimensional tensor through variational mode decomposition (VMD) to obtain multiple sub-modes, each focusing on different frequency components in the signal; Virtual Mode Decomposition (VMD) is a signal decomposition and estimation method that adaptively decomposes a time series into multiple Intrinsic Mode Functions (IMFs) with different center frequencies and residuals. In obtaining the IMFs, iteratively searching for the optimal solution of a variational model determines the frequency center and bandwidth of each IMF, thus adaptively achieving effective IMF classification. VMD primarily utilizes Vienna filtering, Hilbert transform, and Fourier transform to achieve mode decomposition. After decomposition, the original signal equals the sum of each mode and the residuals.

[0022] As the number of sequence decompositions increases, MFFNet's prediction performance improves across different prediction time intervals. Benefiting from its parallel network structure, increasing the number of subsequences does not significantly increase model training time costs, but it does increase the hardware resource requirements during training. This invention defaults to setting the number of IMFs derived from VMD to 5.

[0023] The optimization objective of VMD is expressed as: ; Where k represents the mode index, K represents the number of modes obtained from the decomposition, X(t) represents the original signal, j represents the imaginary unit, and u j (t) represents the j-th mode function, α represents the equilibrium factor, and δu k δt represents the change in the modal signal, and δt represents the time step.

[0024] By analyzing each modality, the model can identify key spatiotemporal features that better reflect the dynamic trends of wind speed changes. The VMD decomposition process enables each sub-modality to more clearly capture local features related to wind speed changes, helping subsequent prediction modules to handle wind speed variations more accurately. Simultaneously, this process effectively reduces interference from data noise and non-stationarity, ensuring the reliability of the extracted features.

[0025] Step 4: Input the decomposed submodalities into the multi-feature fusion network and use the MFFNet network to extract local spatiotemporal features; MFFNet consists of two parts: a residual network and a feature fusion network. The residual network includes one convolutional module and six residual modules, which compress the features twice to obtain three scales. A 3×3 convolutional kernel can aggregate wind speed data from nine wind turbines within a 3×3 region in a single operation, capturing local spatiotemporal dependencies. The feature fusion network merges shallow, medium, and deep features. To eliminate feature scale differences, bilinear interpolation is first used to restore the resolution to match the input, and then the data is stitched together along the channel direction to achieve a unified representation of multi-layer information. Through this design, MFFNet can simultaneously focus on global trends and preserve local details, effectively mitigating the interference caused by differences in wind speed variations between edge wind turbines and neighboring units, thereby improving the robustness and accuracy of predictions.

[0026] Step 5: Calculate the global attention matrix through a multi-head attention mechanism, and combine it with residual connections and normalization modules to enhance feature learning capabilities; Step 6: Predict the extracted features using a linear layer and a softmax activation function, and output the normalized prediction power for each station; Step 7: Through inverse normalization, the prediction results are transformed into the actual wind speed prediction value at the target time point.

[0027] In this embodiment, step 5 first reshapes the local spatiotemporal features extracted by MFFNet, converting them into the form of h×w, hidden_dim. This reshaping uses a row-major approach, expanding the matrix into a one-dimensional form to form the global attention matrix. The expression is: ; Where I is the global attention matrix and Hn is the local spatiotemporal feature.

[0028] The global attention matrix is ​​calculated using a multi-head attention mechanism (MHA). The calculation process of the multi-head attention mechanism is as follows: ; Where Q represents the query matrix, K represents the key matrix, V represents the value matrix, T represents the matrix transpose operation, and dk represents the dimension of the key vector.

[0029] In the output phase, MFFNet ensures the flow and stability of information through layer normalization and residual connection operations. Assuming the output is O, it can be represented as: ; Here, H represents the output of the previous layer, and I represents the model input.

[0030] Ultimately, the model output has a dimension of (h×w, hidden_dim). The hidden layer dimension is reduced to 1 using a linear layer, and then passed through a softmax activation function to obtain the normalized predicted power for each station. Finally, an inverse normalization operation transforms the prediction results into actual wind speed predictions for the target time point.

[0031] The proposed Multi-Feature Fusion Network (MFFNet) is compared with other networks, showing the mean absolute error (MAE), mean squared error (MSE), and symmetric mean absolute percentage error (SMAPE) results of different methods over 12 months, as shown in the table below.

[0032]

[0033] This invention proposes an ultra-short-term wind speed prediction method based on matrix spatiotemporal eigenvalue decomposition and multi-feature fusion to achieve overall prediction of large-scale wind farms. Ablation experiments and comparative experiments show that the introduction of shallow and mid-layer features effectively improves the model's prediction performance, enabling the model to focus on wind speed variation trends within different areas of the wind farm and mitigating the negative impact of varying wind speed trends among adjacent wind turbines on the model's prediction accuracy. The proposed ultra-short-term wind speed prediction method based on matrix spatiotemporal eigenvalue decomposition and multi-feature fusion demonstrates reliable prediction performance throughout the year and in all months. Furthermore, its predictive advantages become more pronounced with increasing prediction time intervals, exhibiting superior performance.

[0034] The MFFNet proposed in this invention fuses shallow, mid-level, and deep features, compensating for lost local information while focusing on global information, thereby improving the model's prediction accuracy. Shallow features mainly originate from the spatiotemporal features of the input sub-matrix, typically including local information from various sites within a wind farm. Parallel sub-networks in MFFNet focus on different local information within these sub-matrix spatiotemporal features, capturing the inherent patterns within these local spatiotemporal features. Mid-level features are global features further fused and enhanced based on shallow features through a multi-head self-attention mechanism. Deep features originate from the fusion results of multiple parallel sub-networks. The MFFNet result gradually extracts deeper, more abstract spatiotemporal dependencies from shallow and mid-level features through multi-layer stacking. Features from each layer are fused through residual connections and layer normalization.

[0035] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for ultra-short-term wind speed prediction based on matrix spatiotemporal eigenvalue decomposition and multi-feature fusion network, characterized in that, Includes the following steps: Step 1: Obtain real-time data for each wind turbine in the wind farm, and map the data into a two-dimensional matrix based on geographical location information to form matrix spatial features; Step 2: Superimpose the matrix space features at consecutive time points to form a three-dimensional tensor containing time information in the channel direction; Step 3: Decompose the three-dimensional tensor by variational mode decomposition to obtain multiple sub-modes, each sub-mode focusing on different frequency components in the signal; Step 4: Input the decomposed submodalities into the multi-feature fusion network and use the MFFNet network to extract local spatiotemporal features; Step 5: Calculate the global attention matrix through a multi-head attention mechanism, and combine it with residual connections and normalization modules to enhance feature learning capabilities; Step 6: Predict the extracted features using a linear layer and a softmax activation function, and output the normalized prediction power; Step 7: Through inverse normalization, the prediction results are transformed into the actual wind speed prediction value at the target time point.

2. The prediction method according to claim 1, characterized in that, The process of constructing the matrix space features in step 1 includes: The real-time data of the wind turbine is mapped into a two-dimensional matrix based on its latitude and longitude coordinates. Each element of the matrix represents the state of the wind turbine at a specific location at a certain moment.

3. The prediction method according to claim 1, characterized in that, The three-dimensional tensor has dimensions of Channel × Height × Width, where Channel represents the number of channels and different feature dimensions; Height and Width represent the number of wind turbines in the region to be determined, respectively, in the vertical and horizontal directions, and are expressed as follows: ； Among them, X n (t) represents the characteristics of the nth wind turbine at time t.

4. The prediction method according to claim 1, characterized in that, The variational mode decomposition process in step 3 includes: The input three-dimensional tensor is decomposed into multiple sub-modes by using variational optimization methods. Each sub-mode has an independent frequency center and bandwidth constraint. An optimization algorithm using the alternating direction multiplier method is employed to update the bandwidth and center frequency of each mode in the frequency domain, ensuring the accuracy of the decomposition. The optimization objective of variational mode decomposition is expressed as: ； Where k represents the mode index, different modes represent different frequencies after wind energy information decomposition, K represents the number of modes obtained from the decomposition, X(t) represents the original three-dimensional signal, j represents the imaginary unit, and u j (t) represents the j-th mode function, α represents the equilibrium factor, and δu k δt represents the change in the modal signal, and δt represents the time step.

5. The prediction method according to claim 1, characterized in that, In step 5, the local spatiotemporal features extracted by MFFNet are first reshaped and converted into the form of (h×w, hidden_dim).

6. The prediction method according to claim 1, characterized in that, In step 5, the global attention matrix is ​​calculated using a multi-head attention mechanism. The calculation process of the multi-head attention mechanism is as follows: ； Where Q represents the query matrix, indicating the prediction demand; K represents the key matrix, indicating the spatiotemporal feature index of historical wind farms; V represents the value matrix, i.e., wind speed feature values; T represents the matrix transpose operation; and d k This represents the dimension of the key vector.

Images