Wind power prediction method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of wind power prediction, in particular to a wind power prediction method and device, electronic equipment and a storage medium, aiming at power grid risks caused by wind power randomness and volatility, high-precision wind power prediction is the key to guarantee the safety and energy efficiency of a power grid; aiming at the problems of data mode aliasing, feature screening limitation, insufficient seasonal adaptation and the like of an existing method, the method comprises the following steps of S1, extracting an original photovoltaic power generation time sequence and mining deep information based on dual data decomposition of CEEMDAN-PSO-VMD; s2, photovoltaic power prediction modeling based on Transform-TCN-SVR: carrying out independent modeling on the sub-components of the specific frequency domain by using the Transform, the TCN and the SVR, and obtaining prediction results of the sub-components; and S3, prediction result integration based on an Attention mechanism, wherein the prediction result integration is used for identifying a non-linear relationship among the sub-components and obtaining a final prediction result.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and more specifically, to wind power prediction methods, devices, electronic equipment, and storage media. Background Technology

[0002] With the intensification of global climate change, countries are accelerating their energy transition to reduce dependence on fossil fuels. Photovoltaic energy, with its advantages of being clean, renewable, and widely distributed, has become an important direction for new energy development. However, the inherent intermittency and volatility of photovoltaic power generation pose a serious challenge to the stable operation of the power grid. Accurate photovoltaic power forecasting is one of the key technologies for ensuring the safe and stable operation of the power system.

[0003] Current photovoltaic (PV) power prediction methods are mainly divided into two categories: physical modeling and data-driven methods. Physical models predict power by establishing mathematical relationships between solar irradiance, PV cell characteristics, and environmental conditions, typically relying on numerical weather prediction (NWP) and satellite data. Although physical models have clear physical meaning, their complex mathematical derivations and high computational costs limit their practical application. Data-driven methods predict power by mining potential patterns in historical data, including statistical and artificial intelligence methods. Traditional statistical methods, such as exponential smoothing and autoregressive integrated moving average (ARIMA), are simple and easy to use, but their linear assumptions make it difficult to characterize the complex nonlinear characteristics of PV power, resulting in limited prediction accuracy when faced with drastic fluctuations.

[0004] In recent years, deep learning has made groundbreaking progress in the field of time series forecasting due to its powerful nonlinear modeling capabilities. Models such as Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Transformers have been widely applied to photovoltaic forecasting. For example, Huang et al. constructed a CNN-based forecasting model that effectively extracted multidimensional meteorological features; Qing et al. demonstrated that LSTM can improve the prediction accuracy by 42.9% compared to traditional BP networks.

[0005] Despite the significant achievements of deep learning methods, their direct application to photovoltaic (PV) forecasting still faces the challenge of insufficient accuracy. To address this, researchers have proposed a forecasting framework based on signal decomposition. This framework utilizes methods such as Discrete Wavelet Transform (DWT), Empirical Mode Decomposition (EMD), and Variational Mode Decomposition (VMD) to decompose complex signals into multiple simpler components. Liu et al. used wavelet packet decomposition to reduce the forecasting difficulty and summed the prediction results of the sub-signals to obtain the final result; Zhai et al. demonstrated that the VMD-Transformer framework can effectively improve forecasting accuracy. The "divide and conquer" strategy—predicting the decomposed sub-signals separately and combining the prediction results of each sub-signal to obtain the final prediction result—has become a mainstream approach for improving PV forecasting performance.

[0006] In summary, although numerous studies have demonstrated that employing a data decomposition framework can effectively improve photovoltaic power prediction results, the following key issues still need to be addressed in existing research: (1) Insufficient single decomposition: The high non-stationarity of photovoltaic power means that the components after a single decomposition still contain complex patterns, especially the high-frequency components, which still fluctuate violently, affecting the accuracy of subsequent modeling. The inherent limitations of a single decomposition algorithm also make it difficult to fully extract multi-scale features.

[0007] (2) The modeling strategy is too simplistic: Existing studies often use the same model to process all components or simple integration, ignoring the differences in characteristics of components at different frequencies. In fact, the linear characteristics of low-frequency components are suitable for lightweight models, while using complex models not only increases the computational burden, but may also lead to overfitting.

[0008] (3) The fusion method is too simplistic: Most studies fuse the prediction results of each component by direct addition or linear regression, ignoring the influence of the decomposition residual term. This simplification introduces systematic nonlinear error, which limits the improvement of the final prediction accuracy.

[0009] To address the aforementioned challenges, this invention proposes a wind power prediction method, device, electronic equipment, and storage medium. Summary of the Invention

[0010] The purpose of this invention is to address the key problems in existing photovoltaic power prediction methods, such as insufficient single-step decomposition (the high non-stationarity of photovoltaic power means that the components after a single decomposition still contain complex patterns, especially high-frequency components which still exhibit drastic fluctuations, and the inherent limitations of a single decomposition algorithm make it difficult to fully extract multi-scale features), simplistic modeling strategies (often using the same model to process all components or simple integration, ignoring the characteristic differences of different frequency components, and using complex models for low-frequency components can easily increase the computational burden and lead to overfitting), and overly simplified fusion methods (mostly fusing the prediction results of each component by direct addition or linear regression, ignoring the influence of the decomposition residual term, and introducing systematic nonlinear errors). This invention aims to provide a high-precision, high-stability photovoltaic power prediction method that is adaptable to different seasons and data characteristics, ensuring the safe and stable operation of the power system.

[0011] To achieve the above objectives, the present invention provides a wind power prediction method, comprising the following steps: S1. The CEEMDAN-PSO-VMD dual data decomposition strategy is adopted. Specifically, the original photovoltaic power generation time series is initially decomposed to obtain multiple intrinsic mode function sub-components. Four types of indices, namely sample entropy, approximate entropy, permutation entropy and spectral entropy, are introduced to quantify the complexity of each intrinsic mode function sub-component. Based on the results of complex quantification, all sub-components are divided into three categories: high frequency, medium frequency, and low frequency through spectral clustering. Sub-components of the same category are summarized and reconstructed to obtain high frequency, medium frequency, and low frequency reconstructed sub-components. By using particle swarm optimization algorithm and optimizing the decomposition modulus and equilibrium parameters of variational mode decomposition, the optimized variational mode decomposition algorithm is applied to the reconstructed high-frequency sub-components to perform secondary decomposition, and finally obtains the high-frequency sub-components, mid-frequency reconstructed sub-components, and low-frequency reconstructed sub-components after secondary decomposition. S2. Photovoltaic power prediction modeling based on Transformer-TCN-SVR: For the reconstructed sub-components in different frequency domains in step S1, Transformer, TCN, and SVR models are used to perform separate modeling to obtain the prediction results for each sub-component. Prediction results and prediction results ; S3. Attention-based prediction result integration: The attention mechanism is used to capture the degree of nonlinear correlation between each sub-component, and the degree of correlation is used to integrate the prediction results. Prediction results and prediction results Dynamic weights are assigned, and weighted fusion is performed to ultimately output the predicted photovoltaic power. .

[0012] As a further improvement to this technical solution: the CEEMDAN-PSO-VMD dual data decomposition strategy in step S1 specifically includes the following steps: S1.1: Collect the original photovoltaic power generation time series, and use CEEMDAN to decompose the original photovoltaic power generation time series to obtain multiple intrinsic mode function sub-components. ; S1.2: Introduce four entropy metrics to evaluate the complexity of each component. The entropy metrics include sample entropy, approximate entropy, permutation entropy, and spectral entropy. S1.3: Based on the entropy index of each sub-component, the sub-components of multiple intrinsic mode function sub-components are clustered into three sub-clusters in different frequency domains through spectral clustering; the first category is the high-frequency sub-component, the second category is the mid-frequency sub-component, and the third category is the low-frequency sub-component. S1.4: Summarize the sub-components of the same cluster to obtain the reconstructed high-frequency, mid-frequency, and low-frequency sub-components, denoted as S. h S m and S l ; S1.5: The high-frequency sub-components of the secondary decomposition are obtained by further decomposing and reconstructing them using the VMD algorithm optimized by PSO. .

[0013] As a further improvement to this technical solution, in step S1.1, CEEMDAN is an improved version of the EMD and EEMD algorithms. CEEMDAN can adaptively divide the original time series into multiple intrinsic mode function sub-components and a residual. Specifically: adaptive noise is added to the k-th order residual. Add Gaussian white noise and standard deviation Generate N sets of noisy signals, the specific expression is: ; Decompose the noisy signal: Decompose each group of noisy signals Perform EMD decomposition and extract the first intrinsic mode function sub-component. The specific expression is as follows: ; Calculate the average IMF and extract the residuals: Take the average of the intrinsic mode function subcomponents of the Nth decomposition to obtain the (k+1)th order intrinsic mode function subcomponent, and update the residuals: ; ; Repeat the above steps until the residual is found. No longer containing local extrema, the final decomposition result is obtained:

[0014] Where K is the total number of intrinsic mode function subcomponents. For the final residual; The repeated steps are to add adaptive noise, decompose the noisy signal, and calculate the average intrinsic mode function sub-components and extract the residuals.

[0015] As a further improvement to this technical solution, in step S1.5, the VMD algorithm determines the center frequency and bandwidth of each mode by constructing a variational model and iteratively solving it. Specifically: Parameter initialization: Set the number of decomposed modes K and the initial center frequency. Balance parameters Lagrange multipliers Number of iterations and convergence threshold ; The modal functions and center frequencies are updated using the following expressions: ; ; Among them, Lagrange multipliers Subject to balance parameters The impact; Determine convergence: Calculate residual energy Specifically, it is expressed as follows: ; If residual energy < If the iteration stops, then the mode function and center frequency are updated, and the update continues; finally, the decomposition result and residual are obtained.

[0016] As a further improvement to this technical solution, step S1 uses a particle swarm optimization algorithm to optimize the decomposition modulus K and the equilibrium parameters. Choice: Particle Swarm Optimization (PSO) is a global optimization algorithm proposed by Kennedy and Eberhart based on simulations of bird predation. PSO searches for the optimal solution among random particles iteratively. Specifically, each solution is treated as a particle in d-dimensional space, possessing two attributes: velocity and position. Suppose that in d-dimensional space, the particle population is M, and the position of particle i is... Corresponding The formulas for the changes in velocity and position of particle i in the d-dimensional components of the search space are as follows: ; ; in, This represents the velocity of the i-th particle in the t-th iteration; represents the position of the i-th particle in the t-th iteration; w is the inertia weight, used to control the degree to which the particle retains its original velocity; c1 and c2 represent the cognitive learning factor and the social learning factor, respectively; r1 and r2 are random numbers between [0,1]. This represents the best position in the history of the i-th particle; This is the best position in the entire history of the particle swarm.

[0017] As a further improvement to this technical solution, step S2 includes the following process: Step 1: Step S2 uses the random forest algorithm to screen the key influencing factors of each sub-component to assess the stability of the prediction results of each sub-component; Step 2: Based on the characteristics of different frequency band sub-components and related prediction models, the Transformer model is selected to predict high-frequency components, the TCN model is selected to predict mid-frequency components, and the SVR model is selected to predict low-frequency components. Finally, a joint prediction model of Transformer-TC is established.

[0018] As a further improvement to this technical solution, the Transformer model in step S2 includes a data embedding layer, an encoder, and a decoder; The data embedding layer requires a special embedding method adapted to time series data. Data embedding is completed using a linear layer or a one-dimensional convolutional layer, and then combined with position encoding before being input to the subsequent encoder and decoder. The encoder and decoder are both composed of a self-attention mechanism and a feedforward neural network.

[0019] As a further improvement to this technical solution, the self-attention in step S2 is as follows: ; Where Q is the query matrix, K is the key matrix, and V is the value matrix. To scale the coefficients of the weight matrix, softmax transforms the obtained attention scores into a probability distribution.

[0020] As a further improvement to this technical solution, the TCN in step S2 is a temporal convolutional network, which consists of a series of dilated causal convolutions and residual connections. The receptive field calculated by the dilated convolution becomes: ; Where n represents the number of hidden layers, k represents the kernel size, and dn represents the dilation factor of the nth hidden layer; Causal convolution: If the time series data is and filters Then the causal convolution at x is: ; Residual link: ; Where y is the residual output, x is the residual input, f(x) is the linear transformation result, Conv1D(x) is the one-dimensional convolution operation, and h is the activation function.

[0021] As a further improvement to this technical solution, the SVR in step S2 is least squares support vector regression. By modifying the loss function of SVM, the inequality constraints of SVM are transformed into equality constraints, and the solution is obtained through a system of linear equations instead of quadratic programming. The specific optimization objective function and constraint function of LSSVR are as follows: ; ; Where w is the weight vector, b is the bias term, and e is the weight vector. i For error variables, The regularization coefficient is . A function that maps input to a high-dimensional space. For training purposes; The above formula is solved using the Lagrange function, specifically: ; in, They are Lagrange multipliers; By taking the partial derivative of the Lagrange function and setting it to zero, the optimization problem can be transformed into a system of linear equations: ; in, K is the kernel function, I is a vector consisting entirely of 1s, and I is the identity matrix; Finally, solving the linear equation yields the regression function for LSSVR: .

[0022] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the VMD algorithm optimized based on the PSO algorithm in this invention; Figure 2 This is a schematic diagram of the Informer model structure of the present invention; Figure 3 This is a schematic diagram of the TCN of the present invention; Figure 4 This is a schematic diagram of the initial decomposition results of CEEMDAN in this invention; Figure 5 This is a schematic diagram of the spectral clustering results in this invention; Figure 6 This is a schematic diagram of the spectral clustering results in this invention; Figure 7 This is a schematic diagram of the secondary decomposition results of the high-frequency components in this invention; Figure 8 This is a schematic diagram illustrating the characteristic importance of each sub-component in this invention; Figure 9 This is a schematic diagram showing the prediction results of the model proposed in this invention and the comparative model; Figure 10 , 11 This is a schematic diagram illustrating the predictive performance of the proposed model and the comparative model in this invention; Figure 12 This is a schematic diagram illustrating the prediction performance of all models in this invention; Figure 13 This diagram illustrates the secondary decomposition method of PSO-VMD in this invention across all evaluation metrics. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0025] This invention proposes a hybrid feature selection-seasonal adaptive decomposition-xLSTM-Informer prediction framework, and the specific wind power prediction method includes the following steps: S1. A dual data decomposition based on CEEMD-PSO-VMD is used to address the problem that complex fluctuations still exist in the sub-components after the initial decomposition, affecting prediction accuracy, and that a single decomposition method is insufficient to fully extract the original photovoltaic power generation time series and lacks in-depth information mining. The main steps include: S1.1: Collect the original photovoltaic power generation time series, and use CEEMDAN to decompose the original photovoltaic power generation time series to obtain sub-components. ; S1.2: Introduce four entropy metrics to evaluate the complexity of each component. The entropy metrics include sample entropy (SampEn), approximate entropy (ApEn), permutation entropy (PermEn), and spectral entropy (SpecEn). S1.3: Based on the entropy index of each sub-component, each sub-component is clustered into three sub-clusters in different frequency domains through spectral clustering; the first category is the high-frequency sub-component, the second category is the mid-frequency sub-component, and the third category is the low-frequency sub-component. S1.4: Summarize the sub-components of the same cluster to obtain the reconstructed high-frequency, mid-frequency, and low-frequency sub-components, denoted as Sh, Sm, and Sl.

[0026] S1.5: The high-frequency sub-components of the secondary decomposition are obtained by further decomposing and reconstructing them using the VMD algorithm optimized by PSO. This is used to address the problem of complex fluctuations still existing in the sub-components and to mine the deep information of the original photovoltaic power generation time series; In step S1.1 above, CEEMDAN is an improved version of the EMD and EEMD algorithms. It improves upon the mode confusion problem of the EMD method and the noise residual problem of EEMD. CEEMDAN can adaptively divide the original time series into several intrinsic mode functions (IMF) of the original signal and a residual. Specifically: Add adaptive noise: in the k-th order residual Add Gaussian white noise and standard deviation Generate N sets of noisy signals, the specific expression is: ; Decompose the noisy signal: Decompose each group of noisy signals Perform EMD decomposition and extract the first IMF component. The specific expression is as follows: ; Calculate the average IMF and extract the residuals: Take the average of the IMFs from the Nth decomposition to obtain the (k+1)th order IMF, and update the residuals: ; ; Repeat the above steps (add adaptive noise, decompose the noisy signal, and calculate the average IMF and extract the residuals) until the residuals are obtained. The signal decomposition result is obtained when the signal no longer contains local extrema (i.e., it cannot be further decomposed). ; Where K is the total number of IMFs, This is the final residual.

[0027] The specific working principle of the VMD (Variational Mode Decomposition) algorithm in step S1.5 above is as follows: VMD is a non-recursive signal decomposition method. It determines the center frequency and bandwidth of each mode by constructing a variational model and solving iteratively. Compared with the traditional EMD, it can effectively suppress mode confusion, and has strong robustness to noise and low decomposition complexity. Its implementation process is as follows: Parameter initialization: Set the number of decomposed modes K and the initial center frequency. Balance parameters Lagrange multipliers Number of iterations and convergence threshold .

[0028] The modal functions and center frequencies are updated using the following expressions: ; ; Among them, Lagrange multipliers Subject to balance parameters The impact.

[0029] Determine convergence: Calculate residual energy Specifically, it is expressed as follows: ; If residual energy < If the iteration stops, then return to the second step (update the mode function and center frequency) and continue updating; finally, obtain the signal decomposition result and residual.

[0030] Furthermore, as can be seen from the relevant formulas of VMD mentioned above, if the VMD decomposition algorithm is used, the original parameters need to be set, and its decomposition effect depends on the decomposition modulus K and the balancing parameter. The settings, if the decomposition modulus K and the balancing parameters are... Improper settings can have a significant impact on VMD decomposition results. Therefore, this paper uses the Particle Swarm Optimization (PSO) algorithm to optimize the decomposition modulus K and equilibrium parameters. The specific choices are: The aforementioned PSO optimization algorithm is a global optimization algorithm proposed by Kennedy and Eberhart based on the simulation of bird predation. The PSO optimization algorithm searches for the optimal result among random particles through an iterative process. Specifically, each solution is treated as a particle in d-dimensional space, possessing two attributes: velocity and position. Its expression is as follows: Suppose that in d-dimensional space, the particle population is M, and the position of particle i is... Corresponding Then the d-dimensional components of particle i in the search space The formulas for the changes in velocity and position in the middle are as follows: ; ; in, This represents the velocity of the i-th particle in the t-th iteration; represents the position of the i-th particle in the t-th iteration; w is the inertia weight, used to control the degree to which the particle retains its original velocity; c1 and c2 represent the cognitive learning factor and the social learning factor, respectively; r1 and r2 are random numbers between [0,1]. This represents the best position in the history of the i-th particle; This is the best position in the entire history of the particle swarm; The advantages of the PSO algorithm described above are its simplicity, few parameters, and fast convergence speed; by adjusting the inertia weight w and learning factors c1 and c2, the algorithm's global exploration and local exploitation capabilities can be balanced; finally, the VMD decomposition process based on PSO optimization is as follows: Figure 1 As shown.

[0031] S2. Photovoltaic power prediction modeling based on Transformer-TCN-SVR: Step S2 in this work utilizes Transformer, TCN, and SVR to model individual sub-components in a specific frequency domain and obtains the prediction results for each sub-component. , and It mainly includes the following processes.

[0032] Step 1: To improve the accuracy of photovoltaic power prediction, step S2 uses the random forest algorithm to screen the key influencing factors of each sub-component and the stability of the prediction results of each sub-component.

[0033] Step 2: Based on the characteristics of different frequency band sub-components and related prediction models, the Transformer model is selected to predict high-frequency components, the TCN model is selected to predict mid-frequency components, and the SVR model is selected to predict low-frequency components. Finally, a Transformer-TC joint prediction model is established to effectively solve the problem that a single prediction model is difficult to adapt to sub-signals with different frequency characteristics. The Transformer model in step S2 was first applied to the field of natural language processing and achieved excellent performance. Thanks to its self-attention mechanism and long-term memory capabilities derived from positional encoding, it has also been widely applied to time series prediction in recent years. More specifically, the Transformer mainly consists of three parts: a data embedding layer, an encoder, and a decoder, as shown in the diagram. Figure 1 As shown, where: Data embedding layer: Unlike language models, time series data requires special embedding methods. Unlike language models, which commonly embed each word into random points in d-dimensional space, time series data typically uses linear layers or convolutional layers to embed the data and combine it with positional encoding before inputting it into the subsequent encoder and decoder.

[0034] Encoder and Decoder: The encoder transforms the input data into high-dimensional features and integrates contextual information, while the decoder reconstructs these features. Each encoder and decoder consists of two core modules: a self-attention mechanism and a feedforward neural network. Unlike traditional recurrent neural networks (RNNs) and convolutional neural networks (CNNs), the Transformer utilizes a self-attention mechanism to capture the dependencies between elements in a sequence, making it inherently parallelizable and more efficient at modeling long-range dependencies. The self-attention mechanism assigns weights to the input sequence by calculating attention scores and obtains the interaction patterns of different parts of the input sequence, as shown below: ; Where Q is the query matrix, K is the key matrix, and V is the value matrix. To scale the coefficients of the weight matrix, softmax transforms the obtained attention scores into a probability distribution.

[0035] To enhance the expressive power of the model and improve the diversity of feature extraction, a multi-head self-attention mechanism is usually adopted in practical modeling, and its expression is as follows: ; The calculation for each head is as follows: ; in, , , , Let be a learnable matrix, and headi represent the data of the i-th attention head; The output of multi-head attention is fed into the feedforward neural network, and each sub-layer is connected by residual links and normalization layers, as shown in the following equation: ; Where x is the original data input, The output obtained through nonlinear transformation, This is the normalized layer.

[0036] In step S2, TCN stands for Temporal Convolutional Network, which is specifically a time series prediction model based on convolutional neural networks, such as... Figure 1 As shown, TCN consists of a series of dilated causal convolutions and residual connections; its special network structure not only has a flexible receptive field, but also avoids the gradient vanishing and gradient exploding problems caused by RNN networks. More details: Dilated Convolution: Because traditional convolutional neural networks are easily limited by the size of the convolutional kernel and the depth of the network structure when modeling time series, in a standard n-layer convolutional network, if the kernel size is k, the calculated receptive field is as follows: ; TCN, by introducing a dilated convolution mechanism, changes the relationship between the receptive field and the hidden layer of the neural network from a linear relationship to an exponential relationship, and its receptive field becomes: ; Where n represents the number of hidden layers, k represents the kernel size, and dn represents the dilation factor of the nth hidden layer.

[0037] Causal convolution: Unlike traditional convolutional networks, causal convolution ensures that future data does not leak into past data. Specifically, if the time series data is... and filters Then the causal convolution at x is: ; Residual Links: Similar to residual links in Transformers, the residual linking mechanism aids in the network training process and effectively mitigates the gradient explosion and vanishing problems that occur when networks are too deep. Through residual links and TCNs, historical information can be effectively conveyed, thereby improving the modeling ability of time series data. In TCNs, residual links can be represented as:

[0038] Where y is the residual output, x is the residual input, f(x) is the linear transformation result, Conv1D(x) is the one-dimensional convolution operation, and h is the activation function; In step S2, SVR stands for Least Squares Support Vector Regression. Specifically, Least Squares Support Vector Regression was proposed by Suykenns and Vandewalle. It primarily modifies the loss function of SVM, transforming the inequality constraints of SVM into equality constraints. Furthermore, it can be solved using a system of linear equations instead of quadratic programming, significantly reducing computational complexity. The specific optimization objective function and constraint functions of LSSVR are as follows: ; ; Where w is the weight vector, b is the bias term, and ei is the error variable. The regularization coefficient is . A function that maps input to a high-dimensional space. For training purposes; The above formula can be solved using the Lagrange function, as shown below: ; in, It is a Lagrange multiplier.

[0039] By taking the partial derivative of the Lagrange function and setting it to zero, the optimization problem can be transformed into a system of linear equations: ; in, K is the kernel function, 1 is a vector consisting entirely of 1s, and I is the identity matrix.

[0040] Finally, solving the linear equation yields the regression function for LSSVR: .

[0041] S3. Integration of Prediction Results Based on Attention Mechanism: Addressing the issue of systematic nonlinear errors in the final prediction results caused by neglecting residual terms in existing research, the Attention mechanism in step S3 is used to identify the nonlinear relationships between the sub-components and obtain the final prediction results. Among them, the prediction results , and The prediction results of the sub-components in step S2 (e.g., trend components, smoothing components, residual components, etc. that can represent the sequence); The attention function is used to assign different weights to each sub-component (the weights are determined by the degree of nonlinear correlation between the sub-components), and then weighted and fused to obtain the final prediction value.

[0042] Taking scaled dot product attention (suitable for sub-component fusion scenarios) as an example, the specific calculation logic of the Attention mechanism is as follows: Calculate attention weights: for each sub-component Calculate its similarity to other sub-components (i.e., attention score), and then obtain the weights through Softmax normalization: ; Where d is the dimension of the sub-component (used to scale the score and avoid gradient explosion).

[0043] Weighted fusion of sub-components: All sub-components are weighted and summed using attention weights to obtain the final prediction result. ; The data in steps S1, S2, and S3 above comes from a photovoltaic power generation device in Fuzhou, Fujian Province, China, from January 3, 2022 to April 30, 2023. The data includes various photovoltaic power generation related information, such as power generation capacity, weather information, and the device's latitude and longitude. The photovoltaic power generation device collects data every 15 minutes, 24 hours a day. Considering that the photovoltaic power generation device does not generate power at night, only the relevant data from 7:00 AM to 7:00 PM each day is retained. After data processing, the dataset contains 23,233 data points. This example uses the last month of the dataset (April 2023) as the test set, containing 1,440 data points; the remaining data serves as the training and validation set, totaling 21,793 data points. Descriptive statistics and simplified names of the relevant data are shown in Table 1.

[0044] Table 1 Furthermore, to eliminate the influence of different scales on the training of deep learning models, the data undergoes Min-Max normalization before being input into the deep learning model, the expression of which is as follows: ; in, The original data, and This represents the maximum and minimum values ​​in the dataset. This is the result of normalization.

[0045] All experiments in steps S1, S2, and S3 above were conducted on a 64-bit PC platform equipped with an Intel(R) Core(TM) i5-12600KF CPU@3.70GHz and an NVIDIA GeForce RTX 4060 Ti GPU (16GB). The deep learning models involved were all built and trained based on the TensorFlow (version 2.x) open-source machine learning framework.

[0046] The specific working principle of the CEEMD-PSO-VMD dual decomposition in step S1 is as follows: The initial decomposition and reconstruction based on CEEMDAN first utilizes CEEMDAN to perform an initial decomposition of the original photovoltaic data. The specific decomposition results of each sub-component are as follows: Figure 1 As shown; by Figure 1 It is known that the original photovoltaic power data was decomposed into 15 multi-scale components, and each sub-component exhibited different frequency characteristics. This significantly increased the computational complexity of subsequent secondary decomposition and deep learning prediction. Therefore, four entropy indices—Sample Entropy (SampEn), Approximate Entropy (ApEn), Permutation Entropy (PermEn), and Spectral Entropy (SpecEn)—were introduced to fully quantify the complexity and irregularity of each sub-component. Using the obtained entropy matrix as input, spectral clustering was employed to cluster each component into three clusters in different frequency domains, effectively reducing the complexity of subsequent work. The clustering results of each IMF were visualized using PCA dimensionality reduction, as shown below. Figure 1 As shown; the clustering categories of each sub-component are shown in column 6 of Table 2.

[0047] Table 2 Depend on Figure 1 The mid-spectral clustering results show that, after removing the residual components, the most complex IMF1-IMF3 can be classified as high-frequency sub-components, the relatively complex IMF4-IMF8 can be classified as mid-frequency sub-components, and the least complex IMF9-IMF15 are classified as low-frequency sub-components. Based on this, the initial decomposition results are reconstructed, and the reconstruction results are as follows. Figure 1 As shown; Based on PSO-VMD secondary decomposition: As shown in Table 2, the entropy matrix results indicate that some components still exhibit high complexity after the initial decomposition. This is because the high volatility of the photovoltaic power generation sequence itself makes it difficult for a single decomposition to fully capture its characteristics. A secondary decomposition using VMD is necessary to further extract the deep features of the photovoltaic power generation sequence, given that the high-frequency components still possess significant complexity. However, considering that VMD is heavily influenced by preset parameters, the decomposition modulus K and balance parameters of VMD are optimized with minimizing the envelope entropy as the objective. Optimization was performed; the minimum envelope entropy identified by the PSO optimization algorithm was 0.677, corresponding to a decomposition modulus K of 2 and a balancing parameter. The value is 527.46; the final result of the quadratic decomposition of the high-frequency components under these parameters is as follows: Figure 1 As shown.

[0048] For details, please refer to the following: Figure 1The results show that the high-frequency components after secondary decomposition are more regular, more uniformly distributed, and less complex than the original high-frequency components. This paper will utilize Transformer, TCN, and LSSVR for targeted modeling and prediction based on the mid-frequency and low-frequency sub-components obtained from the initial decomposition and the complexity of the high-frequency sub-components obtained from PSO-VMD, and will employ an Attention mechanism to fit the prediction results. This reduces training costs while improving the accuracy of the final prediction results.

[0049] Step S2 uses the random forest algorithm to screen the key influencing factors of each sub-component and assess the stability of the prediction results for each sub-component. Specifically, the feature factor selection based on random forest is as follows: Each sub-component actually reflects the inherent information of photovoltaic power generation data. For example, low-frequency components can express the trend of photovoltaic data, while high-frequency components correspond to the instantaneous interference of environmental factors on photovoltaic data. This means that the degree of influence of each climate data on each sub-component is different. To improve the accuracy and stability of the prediction results of each sub-component, and considering the non-linear relationship between each climate and photovoltaic data, random forest is used to identify the importance of each climate data to each sub-component, and factors with low importance in each sub-component are eliminated. Finally, the characteristic importance of each sub-component is as follows: Figure 1 As shown; And by Figure 1 It can be seen that for the high-frequency sub-components, Irr (irradiance), RH (relative humidity), Pres (air pressure), Temp (temperature), and Prec (precipitation) are more important, reflecting the rapid interaction and influence of meteorological elements such as atmospheric humidity, pressure, thermal state, and precipitation processes in a short period of time. As an intermediate state between the instantaneous interference and long-term trend of environmental factors on photovoltaic power generation, the mid-frequency sub-component mainly reflects the normal influence of environmental factors on photovoltaic power generation. For traditional methods based on physical modeling, step S2 can predict the photovoltaic power under normal conditions by constructing a mechanism model with irradiance and ambient temperature as inputs. Figure 1 The importance of the mid-frequency subcomponents in the model can fully reflect this point; for the low-frequency subcomponents, the importance of Irr (irradiance intensity) decreases and the importance of Temp (temperature) is the highest, because the long-term thermal effect of photovoltaic equipment plays a key role in the long-term trend of photovoltaic power generation

[35] . In summary, random forests can be used to fully explore the influence of each climate factor on the subcomponents; based on this, the top five factors of importance of each subcomponent are taken as important influencing factors and input into the subsequent prediction models of each subcomponent.

[0050] To verify the superiority of the proposed model, this example selects four commonly used photovoltaic prediction models—LSTM, GRU, TCN, and Transformer—as the basic models for the comparative experiment. Corresponding improvements and innovations are addressed. Each model uses the same training and prediction sets as input, and the parameters of each model are consistent in the same parts. Following the common approach in existing research, each sub-component is predicted by a model with the same architecture, and the final prediction result is obtained through linear addition [7],

[19] ,

[20] ,

[37] . It should be noted that although the SVR model is used to construct the proposed model, its weak nonlinear fitting ability makes it difficult to handle high-frequency sub-components with high complexity. Therefore, this example does not use it as the base model for the comparative experiment. The final model parameters are shown in Table 3.

[0051] Table 3 Evaluation metrics: To assess the accuracy of the predictions and models proposed in this example, root mean square error (RSME), mean absolute error (MAE), and coefficient of determination (R²) are introduced as evaluation metrics. The relevant calculation formulas are as follows:

[0052]

[0053]

[0054] in This represents the actual value of the original data. This represents the model's predicted value. This represents the average value of the original data; Comparative Experiment Results: Based on the above research process, this example introduces four indicators to verify the effectiveness of the proposed prediction framework and prediction model; Table 4 and... Figure 10 The prediction accuracy of the proposed model and the comparison model on the test set is shown; the bold text indicates the best value of the evaluation metric:

[0055] Table 4 Specific Table 4 and Figure 10 The prediction results of the proposed model and the comparison model are presented. It can be seen that the model constructed in this paper has the best fitting ability. The RMSE of the model is 7.999, the MAE is 5.147, and the R2 is 0.971. Compared with the comparison model M3, which has the best prediction effect, the RMSE and MAE of the model constructed in this example are reduced by 0.410 and 0.307, respectively, and the prediction performance is improved by 5% and 6%.

[0056] Figure 10 and Figure 12 To further and more intuitively demonstrate the predictive performance of the proposed model and the comparison model; Figure 10 Line graphs showing the prediction results and actual values ​​of each model are presented, which show that the prediction results of the proposed model are closer to the actual values. Figure 12 Scatter plots showing the relationship between the predicted results and the actual values ​​for each model are presented; the results indicate that the proposed model has the highest goodness of fit and smaller prediction errors; furthermore, a comprehensive analysis... Figure 10 and Figure 12 It can also be found that the decomposition-based framework can effectively improve the prediction performance of the base model, especially the secondary decomposition process can greatly improve the prediction performance of the base model. Compared with the single decomposition process, the secondary decomposition process can generally improve the performance of the base model by more than 20%. This phenomenon initially verifies the problem that the sub-components after the initial decomposition still have complex fluctuations that affect the prediction accuracy, and the effectiveness of the secondary decomposition process in greatly reducing the complexity of the complex sub-components and effectively improving the prediction performance.

[0057] Statistical Testing: To further verify the statistical significance of the performance advantage of the proposed model, the Diebold-Mariano (DM) test was used to analyze the prediction results. This test is a standard method for evaluating whether the performance difference between two prediction models is statistically significant. Using mean squared error (MSE) as the loss function, a one-sided DM test was performed on M13 and each benchmark model. The test results show that the prediction accuracy of the proposed model is significantly better than the comparison models. Detailed test results are shown in Table 5.

[0058] Table 5 To further verify the stability and general applicability of the model, experiments were conducted using the test sets from July 2022, October 2022, and January 2023. The comparison models were the M3, M6, M9, and M12 models after secondary decomposition, as shown in Table 6 below.

[0059] Table 6 The seasonality analysis results fully demonstrate the robustness and adaptability of the method presented in this paper; such as Figure 12As shown in Table 6, the predictive performance of all models exhibits significant seasonal differences: the highest prediction accuracy is observed in winter (lowest RMSE: 5.40), while the lowest accuracy is observed in summer (lowest RMSE: 12.02). This phenomenon reveals the significant impact of seasonal meteorological factors on model performance. Experimental results show that the proposed M13 model maintains optimal performance across all seasons. Compared to the suboptimal model M12, it achieves RMSE improvements of 6.8%, 6.2%, and 4.8% in summer, autumn, and winter, respectively, demonstrating stronger adaptability, particularly in summer when prediction is most challenging. Analysis of the prediction curves reveals that the high sunlight intensity and variable weather conditions in summer (July) lead to drastic data fluctuations and numerous abrupt changes (such as intermittent shading caused by cloud cover), posing a severe challenge to model robustness. In contrast, stable weather conditions and lower sunlight intensity in winter result in data exhibiting regular periodic characteristics, which is beneficial for improving prediction accuracy. The cross-seasonal stability of M13 not only validates the effectiveness of the proposed method but also provides important technical support for developing differentiated seasonal scheduling strategies for photovoltaic power plants.

[0060] Ablation Experiment Results: Through model comparison and generalization analysis, the predictive performance of the proposed method has been preliminarily verified. To further explore the independent contribution of each technical module, this section uses the original test set as the prediction object and designs ablation experiments for the quadratic decomposition strategy, differentiated modeling, and nonlinear fitting mechanism, respectively. Each time, only one module setting is changed, while the configuration of other modules remains unchanged.

[0061] Effectiveness of the quadratic decomposition strategy: To verify the superiority of the PSO-VMD quadratic decomposition strategy, this paper compares it with single-step decomposition and quadratic decomposition schemes employing wavelet decomposition, singular value decomposition, and EMD. The results are shown in Table 7 and... Figure 12 As shown:

[0062] Table 7 Refer to Table 5 and Figure 12 As can be seen, the proposed PSO-VMD quadratic decomposition method achieves optimal or near-optimal levels across all evaluation metrics. Particularly noteworthy is that while the wavelet transform-based M17 method performs similarly to the proposed method in terms of MAE (5.150 vs 5.147), its RMSE is significantly higher (9.730 vs 7.999). This difference indicates that while the M17 method provides accurate predictions in most cases, it lacks stability when handling abrupt changes or outliers in the data, making it prone to significant prediction bias. Figure 12The violin plot of the error distribution shows that the proposed PSO-VMD quadratic method has the most concentrated and symmetrical error distribution, exhibiting ideal normal distribution characteristics; while other methods show varying degrees of dispersion and skewness. The time-series error plot also clearly demonstrates the differences in performance of each method when handling outliers (such as samples 358, 717, and 1075). The proposed method has a significantly smaller error peak and can recover to the normal prediction level more quickly. Furthermore, compared with single decomposition (M14) and various quadratic decomposition methods, all quadratic decomposition strategies outperform single decomposition. The RMSE based on single decomposition is 11.575, while the RMSE based on quadratic decomposition methods is significantly reduced, further demonstrating the necessity of quadratic decomposition.

[0063] In summary, the ablation experiment results fully demonstrate the effectiveness and superiority of the PSO-VMD secondary decomposition strategy, and verify the rationality of the method design presented in this paper.

[0064] Effectiveness of differentiated modeling: To evaluate the advantages of the component-specific prediction strategy proposed in this paper, the prediction performance of the hybrid model scheme and the single base model scheme are compared; Table 8 shows the detailed comparison results:

[0065] Table 8 As shown in Table 8, the targeted modeling strategy proposed in this paper performs well in multiple evaluation indicators. Compared with the single model scheme, the RMSE of this method is reduced by 7.6%, 8.7% and 6.2% respectively compared with GRU, LSTM and TCN, which significantly improves the prediction accuracy. Although the performance of the Transformer single model is slightly better than the method in this paper (RMSE difference of 1.3% and MAE difference of 2.2%). It is worth noting that the strategy in this paper adopts the lightweight LSSVR model in the prediction of low frequency components, but this small difference requires higher computational complexity. Considering the strict requirements of power system for real-time response

[40] , this paper avoids the computational burden caused by using a complex model globally by configuring appropriate models for different components. While ensuring the prediction accuracy, it significantly improves the computational efficiency and is more in line with the needs of engineering applications. In summary, the experimental results fully verify the effectiveness and practicality of the targeted modeling strategy and prove that the design concept of model selection based on component characteristics has important application prospects.

[0066] Effectiveness of Nonlinear Fitting: Based on verifying the effectiveness of the two key modules, quadratic decomposition and targeted modeling, this paper focuses on exploring the contribution of the Attention mechanism to the fusion of component prediction results. Experiments were designed to compare four fusion strategies: simple linear superposition, ridge regression, multilayer perceptron (MLP), and the proposed Attention fusion mechanism. Table 9 shows the performance comparison results of each method.

[0067] Table 9 The experimental results in Table 7 clearly demonstrate the significant impact of different fusion strategies on prediction performance. Among the four fusion methods, the proposed Attention mechanism achieves the best performance with an RMSE of 7.999, which is 7.1%, 3.8%, and 2.0% lower than simple addition, ridge regression, and MLP, respectively. Notably, the nonlinear fusion methods (MLP and Attention) generally outperform the linear methods (simple addition and ridge regression), fully validating the hypothesis that there are complex nonlinear relationships between components. Combined with the aforementioned ablation experiments of quadratic decomposition and targeted modeling, the complete technical framework proposed in this paper—PSO-VMD quadratic decomposition, component-targeted modeling, and Attention fusion—has proven to be an effective way to improve the accuracy of photovoltaic predictions.

[0068] In summary: This paper addresses key issues in photovoltaic power prediction, such as insufficient single-step decomposition, simplistic modeling strategies, and simplistic fusion methods. It proposes an innovative prediction framework based on quadratic decomposition, differentiated modeling, and attention fusion. Through theoretical analysis and experimental verification, the following conclusions are drawn: (1) The multi-level decomposition strategy significantly improves the feature extraction effect; the CEEMDAN-spectral clustering-PSO-VMD secondary decomposition method proposed in this paper effectively solves the limitations of single decomposition. Experiments show that compared with the traditional single decomposition, this method reduces the prediction error by 30.9%, especially for the processing of high-frequency components. The RMSE after secondary decomposition is reduced from 11.575 to 7.999, which fully demonstrates the necessity of deep feature mining.

[0069] (2) The differentiated modeling strategy achieves a balance between accuracy and efficiency. By matching appropriate prediction models (high frequency - Transformer, mid frequency - TCN, low frequency - LSSVR) to different frequency components, the proposed method significantly reduces computational complexity while maintaining accuracy comparable to the best single model (RMSE difference of only 1.3%). This "localized" modeling approach provides a feasible solution for real-time prediction applications.

[0070] (3) Attention fusion mechanism effectively captures nonlinear relationships: Compared with traditional fusion methods such as simple addition and ridge regression, nonlinear fusion based on attention mechanism reduces prediction error by 7.1% and 3.8%, respectively. This mechanism not only improves prediction accuracy, but also provides good interpretability through weight allocation, which helps to understand the contribution of each component to the final prediction.

[0071] (4) The model demonstrates excellent generalization ability and robustness: the seasonal analysis results show that the proposed method maintains optimal performance in different seasons, especially in summer when prediction is most difficult, and can still effectively cope with the challenges brought by high irradiance and variable weather. DM statistical test further confirms the significance of performance improvement.

[0072] 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 preferred examples and are not intended to limit 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 wind power prediction method, characterized in that, Includes the following steps: S1. The CEEMDAN-PSO-VMD dual data decomposition strategy is adopted. Specifically, the original photovoltaic power generation time series is initially decomposed to obtain multiple intrinsic mode function sub-components. Four types of indices, namely sample entropy, approximate entropy, permutation entropy and spectral entropy, are introduced to quantify the complexity of each intrinsic mode function sub-component. Based on the results of complex quantification, all sub-components are divided into three categories: high frequency, medium frequency, and low frequency through spectral clustering. Sub-components of the same category are summarized and reconstructed to obtain high frequency, medium frequency, and low frequency reconstructed sub-components. By using particle swarm optimization algorithm and optimizing the decomposition modulus and equilibrium parameters of variational mode decomposition, the optimized variational mode decomposition algorithm is applied to the reconstructed high-frequency sub-components to perform secondary decomposition, and finally obtains the high-frequency sub-components, mid-frequency reconstructed sub-components, and low-frequency reconstructed sub-components after secondary decomposition. S2. Photovoltaic power prediction modeling based on Transformer-TCN-SVR: For the reconstructed sub-components in different frequency domains in step S1, Transformer, TCN, and SVR models are used to perform separate modeling to obtain the prediction results for each sub-component. Prediction results and prediction results ; S3. Attention-based prediction result integration: The attention mechanism is used to capture the degree of nonlinear correlation between each sub-component, and the degree of correlation is used to integrate the prediction results. Prediction results and prediction results Dynamic weights are assigned, and weighted fusion is performed to ultimately output the predicted photovoltaic power. .

2. The wind power prediction method according to claim 1, characterized in that: The CEEMDAN-PSO-VMD dual data decomposition strategy in step S1 specifically includes the following steps: S1.1: Collect the original photovoltaic power generation time series, and use CEEMDAN to decompose the original photovoltaic power generation time series to obtain multiple intrinsic mode function sub-components. ; S1.2: Introduce four entropy metrics to evaluate the complexity of each component. The entropy metrics include sample entropy, approximate entropy, permutation entropy, and spectral entropy. S1.3: Based on the entropy index of each sub-component, the sub-components of multiple intrinsic mode function sub-components are clustered into three sub-clusters in different frequency domains through spectral clustering; the first category is the high-frequency sub-component, the second category is the mid-frequency sub-component, and the third category is the low-frequency sub-component. S1.4: Summarize the sub-components of the same cluster to obtain the reconstructed high-frequency, mid-frequency, and low-frequency sub-components, denoted as S. h S m and S l ; S1.5: The high-frequency sub-components of the secondary decomposition are obtained by further decomposing and reconstructing them using the VMD algorithm optimized by PSO. .

3. The wind power prediction method according to claim 2, characterized in that: In step S1.1, CEEMDAN is an improved version of the EMD and EEMD algorithms. CEEMDAN can adaptively divide the original time series into multiple intrinsic mode function sub-components and a residual. Specifically, adaptive noise is added to the k-th order residual. Add Gaussian white noise and standard deviation Generate N sets of noisy signals, the specific expression is: ； Decompose the noisy signal: Decompose each group of noisy signals Perform EMD decomposition and extract the first intrinsic mode function sub-component. The specific expression is as follows: ; Calculate the average IMF and extract the residuals: Take the average of the intrinsic mode function subcomponents of the Nth decomposition to obtain the (k+1)th order intrinsic mode function subcomponent, and update the residuals: ； ； Repeat the above steps until the residual is found. No longer containing local extrema, the final decomposition result is obtained: Where K is the total number of intrinsic mode function subcomponents. For the final residual; The repeated steps are to add adaptive noise, decompose the noisy signal, and calculate the average intrinsic mode function sub-components and extract the residuals.

4. The wind power prediction method according to claim 3, characterized in that: In step S1.5, the VMD algorithm determines the center frequency and bandwidth of each mode by constructing a variational model and solving iteratively. Specifically: Parameter initialization: Set the number of decomposed modes K and the initial center frequency. Balance parameters Lagrange multipliers Number of iterations and convergence threshold ; The modal functions and center frequencies are updated using the following expressions: ； ； Among them, Lagrange multipliers Subject to balance parameters The impact; Determine convergence: Calculate residual energy Specifically, it is expressed as follows: ; If residual energy < If the iteration stops, then the mode function and center frequency are updated, and the update continues; finally, the decomposition result and residual are obtained.

5. The wind power prediction method according to claim 4, characterized in that: Step S1 uses a particle swarm optimization algorithm to optimize the decomposition modulus K and the equilibrium parameters. Choice: Particle Swarm Optimization (PSO) is a global optimization algorithm proposed by Kennedy and Eberhart based on simulations of bird predation. PSO searches for the optimal solution among random particles iteratively. Specifically, each solution is treated as a particle in d-dimensional space, possessing two attributes: velocity and position. Suppose that in d-dimensional space, the particle population is M, and the position of particle i is... Corresponding The formulas for the changes in velocity and position of particle i in the d-dimensional components of the search space are as follows: ； ； in, This represents the velocity of the i-th particle in the t-th iteration; represents the position of the i-th particle in the t-th iteration; w is the inertia weight, used to control the degree to which the particle retains its original velocity; c1 and c2 represent the cognitive learning factor and the social learning factor, respectively; r1 and r2 are random numbers between [0,1]. This represents the best position in the history of the i-th particle; This is the best position in the entire history of the particle swarm.

6. The wind power prediction method according to claim 5, characterized in that: Step S2 includes the following process: Step 1: Step S2 uses the random forest algorithm to screen the key influencing factors of each sub-component to assess the stability of the prediction results of each sub-component; Step 2: Based on the characteristics of different frequency band sub-components and related prediction models, the Transformer model is selected to predict high-frequency components, the TCN model is selected to predict mid-frequency components, and the SVR model is selected to predict low-frequency components. Finally, a joint prediction model of Transformer-TC is established.

7. The wind power prediction method according to claim 6, characterized in that: The Transformer model in step S2 includes a data embedding layer, an encoder, and a decoder. The data embedding layer requires a special embedding method adapted to time series data. Data embedding is completed using a linear layer or a one-dimensional convolutional layer, and then combined with position encoding before being input to the subsequent encoder and decoder. The encoder and decoder are both composed of a self-attention mechanism and a feedforward neural network.

8. A wind power prediction device, characterized in that, include: The sub-component analysis module is used to perform initial decomposition of the original photovoltaic power generation time series to obtain multiple intrinsic mode function sub-components; it introduces four types of indices, namely sample entropy, approximate entropy, permutation entropy and spectral entropy, to quantify the complexity of each intrinsic mode function sub-component; Based on the results of complex quantification, all sub-components are divided into three categories: high frequency, medium frequency, and low frequency through spectral clustering. Sub-components of the same category are summarized and reconstructed to obtain high frequency, medium frequency, and low frequency reconstructed sub-components. By using particle swarm optimization algorithm and optimizing the decomposition modulus and equilibrium parameters of variational mode decomposition, the optimized variational mode decomposition algorithm is applied to the reconstructed high-frequency sub-components to perform secondary decomposition, and finally obtains the high-frequency sub-components, mid-frequency reconstructed sub-components, and low-frequency reconstructed sub-components after secondary decomposition. The prediction result output module uses Transformer, TCN, and SVR models to separately model the reconstructed sub-components in different frequency domains and obtain the prediction results for each sub-component. Prediction results and prediction results ; The analysis determines the module, employs an attention mechanism to capture the degree of nonlinear correlation between each sub-component, and uses this correlation degree as the basis for prediction results. Prediction results and prediction results Dynamic weights are assigned, and weighted fusion is performed to ultimately output the predicted photovoltaic power. .

9. An electronic device, characterized in that, include: At least one processor; The system includes a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to cause the at least one processor to perform the wind power prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the wind power prediction method as described in any one of claims 1-7.

Images