A Short-Term Wind Power Prediction Method Based on Adaptive Mode Decomposition and Multi-Scale Temporal Hybrid Network

Abstract

This invention relates to a short-term wind power prediction method based on adaptive mode decomposition and multi-scale time hybrid networks. In wind power data processing, variational mode decomposition (VMD) is used to decompose the original wind power sequence into high-, medium-, and low-frequency mode components with different center frequencies. This solves the problems of non-stationarity and strong fluctuations in wind power sequences, reducing prediction difficulty and improving accuracy. The effectiveness of variational mode decomposition depends on the number of modes and the penalty factor. A slime mold optimization algorithm is used to find the optimal number of modes and penalty factor, solving the problems of traditional parameter settings relying on experience and unstable decomposition results, thus improving the decomposition quality of wind power by VMD. TimeMixer prediction sub-models are established for the high-, medium-, and low-frequency mode components respectively. The TimeMixer model is built based on MLP, which can accurately capture multi-scale fluctuations, adapt to the characteristics of wind power data, and has fast training speed and high efficiency. This method improves the training speed and accuracy of the prediction model, providing strong support for short-term wind power prediction.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation prediction technology, specifically involving a short-term wind power prediction method based on adaptive mode decomposition and multi-scale time feature hybrid modeling. Background Technology

[0002] With the increasing proportion of new energy power generation in the power system, wind power has been widely used due to its advantages such as being clean and renewable. However, wind power is affected by various factors such as wind speed, wind direction, and atmospheric turbulence, exhibiting strong randomness, nonlinearity, and non-stationarity, posing significant challenges to the safe and stable operation and dispatch of the power grid. Therefore, improving the accuracy of short-term wind power forecasting has important engineering application value.

[0003] To mitigate the non-stationarity problem, related studies have proposed using signal decomposition methods such as empirical mode decomposition (EMD) and variational mode decomposition (VMD) to preprocess wind power sequences, decomposing the original sequence into multiple subsequences with different frequency characteristics before modeling and predicting them separately. However, existing mode decomposition methods typically rely on manual experience to set decomposition parameters, resulting in a lack of adaptive parameter selection and poor stability of the decomposition results, thus affecting subsequent prediction performance. Therefore, this paper uses a slime mold optimization algorithm to adaptively select the optimal parameters for variational mode decomposition, which is of great significance for mode decomposition.

[0004] Short-term wind power forecasting is a key technology for improving grid absorption capacity and operational reliability. Existing technologies mainly include statistical methods, traditional machine learning methods, and deep learning methods. Among them, statistical methods have limited ability to characterize nonlinear features; traditional machine learning methods rely on manual feature design and have insufficient generalization ability; although deep learning methods have improved prediction accuracy to some extent, they still have shortcomings in dealing with problems such as multi-timescale aliasing, severe noise interference, and obvious abrupt changes in wind power sequences. On the other hand, for the decomposed subsequences, existing prediction methods mostly use a uniform prediction model structure for modeling, without fully considering the time scale differences between different modes, resulting in insufficient ability of the model to characterize high-frequency fluctuations and low-frequency trends, affecting the overall prediction accuracy. Therefore, this paper proposes a short-term wind power forecasting method that can adaptively determine mode decomposition parameters and construct a matching prediction model structure for different time scale features to improve prediction accuracy and model stability. Summary of the Invention

[0005] This invention aims to provide a short-term wind power prediction method based on adaptive mode decomposition and multi-scale time feature hybrid modeling, addressing the problems in existing wind power prediction technologies, such as mode decomposition parameters relying on experience, insufficient adaptability of prediction models to multi-time-scale features, and the need to improve overall prediction accuracy and stability. This invention introduces an intelligent optimization mechanism to achieve adaptive optimization of mode decomposition parameters and combines a multi-scale time feature hybrid prediction structure to perform targeted modeling of wind power subsequences with different frequency characteristics, thereby effectively improving the accuracy and robustness of short-term wind power prediction.

[0006] The short-term wind power prediction method proposed in this invention, based on adaptive mode decomposition and multi-scale time feature hybrid modeling, consists of the following four steps:

[0007] S1: Adaptively search for key parameters of variational mode decomposition using an intelligent optimization algorithm. The key parameters include the number of components in the mode decomposition and the penalty factor.

[0008] S2: Perform multimodal decomposition on the wind power sequence to obtain several power subsequences with different frequency characteristics. Based on the dominant frequency distribution of the modal components and the sample entropy, divide the power subsequences into modal components of different time scales.

[0009] S3: For different modal components, construct a multi-scale time feature hybrid prediction sub-model that matches their time scale characteristics. Finally, reconstruct the prediction results of each modal component by weighting to obtain the final prediction result of wind power.

[0010] S4: Verify and analyze the feasibility of the proposed prediction method using a simulation platform.

[0011] S1: The performance of Variational Mode Decomposition (VMD) heavily depends on the number of modes K and the penalty factor α. Determining these parameters through experience and trial and error increases workload and does not guarantee optimal results. The slime mold optimization algorithm is a metaheuristic optimization algorithm inspired by the foraging behavior of slime molds in nature. It completely avoids manual parameter selection and achieves adaptive and intelligent selection of VMD parameters. Slime molds form interconnected path networks through oscillations, achieving a balance between exploring food sources (global search) and shrinking paths to concentrate resources on high-quality food (local development). VMD parameter selection is defined as an optimization problem, and the VMD parameters K and α to be optimized are encoded as position vectors of individual slime molds in the algorithm. First, the core parameters of the algorithm are set. Then, the position vectors of slime mold individuals are randomly initialized in the search space. Next, the objective function value corresponding to each individual's position, i.e., the fitness, is calculated. The fitness function design is the most crucial step; its goal is to minimize an index that measures the quality of decomposition. This invention uses envelope entropy as the index, which reflects the sparsity and impact characteristics of the signal. The smaller the entropy, the purer the decomposed modes. Update the weight W and key parameters. W will dynamically change based on the number of iterations and the individual's fitness ranking. Individuals with higher fitness have higher weights, attracting other individuals to gravitate towards them. The formula is as follows:

[0012] (1)

[0013] Where SmellOrder is the individual fitness ranking index. Next, the population position is updated using the following formula:

[0014] (2)

[0015] p is the probability threshold. Indicates a random position. The current optimal position is determined. The fitness of the new position and the old position are compared, and the better position is retained. This process is repeated iteratively until the stopping condition is met.

[0016] S2: Based on S1, when the slime mold optimization algorithm reaches the stopping condition, the globally optimal individual, i.e., the optimal VMD parameter combination, is output. Using this set of optimal parameters, the final VMD decomposition is performed on the original signal to obtain the optimal modal components. VMD is a completely non-recursive signal decomposition method. Its core idea is to adaptively decompose the original signal into K quasi-orthogonal modal functions with specific sparsity. Its core objective is to find a set of modes. and the corresponding center frequency This minimizes the sum of bandwidths for all modes while ensuring accurate reconstruction of the original signal. Modal decomposition is performed on the wind power sequence to obtain subsequences of different frequencies. Then, based on the dominant frequency distribution and sample entropy, the power subsequences are divided into modal components at different time scales. The specific steps are as follows:

[0017] S2.1: First, initialize all modes. Center frequency Lagrange multipliers The value is 0. In the frequency domain, all modes are updated. The update formula for the k-th mode is:

[0018] (3)

[0019] The molecular part is derived from the original signal spectrum. In, subtract all other modes The current estimate, plus the Lagrange multipliers. The denominator is a fraction... The filter is centered at the center frequency. It acts as a bandpass filter; near the center frequency, the denominator is small, and the overall fractional value is large, allowing that frequency component to pass. Far from the center frequency, the denominator increases sharply, the fractional value decreases, and that frequency component is suppressed. The penalty factor α controls the filter's bandwidth; the larger α is, the faster the denominator grows with frequency shift, the narrower the filter, and the smaller the modal bandwidth. Then, all center frequencies are updated. The center frequency is updated by calculating the first moment of the power spectrum of that mode:

[0020] (4)

[0021] Energy distribution of modes Viewed as a density function, the new center frequency is the location of the first moment of this density function. This allows the center frequency of each mode to automatically move to the frequency band where its energy is most concentrated. Furthermore, the Lagrange multipliers need to be updated to strengthen the reconstruction constraints, as shown in the following formula:

[0022] (5)

[0023] This is the update step size, typically taken as the noise tolerance parameter, defaulting to 0 to ensure strict constraint satisfaction. The term in parentheses is the spectrum of the reconstructed residual in the current iteration. Finally, the convergence condition is checked, and k time-domain modal components and their corresponding center frequencies are output.

[0024] S2.2: Following S2.1, k modal components and their corresponding center frequencies are obtained. High frequencies correspond to short-term rapid fluctuations, while low frequencies correspond to long-term slow trends. In addition to the center frequencies, this invention also uses sample entropy to quantify the complexity and randomness of each modal component. Given the embedding dimension m and tolerance r, sample entropy statistically calculates the probability of similarity between vector templates of length m and m+1 in the time series. The larger the sample entropy value, the more complex, irregular, and unpredictable the sequence. Therefore, modal components with high sample entropy can be considered noise, while modal components with low sample entropy are retained. Adaptive modal decomposition separates the originally complex wind power sequence into subsequences with different scale characteristics. Then, based on the dominant frequency distribution and sample entropy, the subsequences are classified to finally obtain effective high-frequency, mid-frequency, and low-frequency modal components.

[0025] After obtaining modal components from the wind power sequence through adaptive mode decomposition, the design steps of the multi-scale time hybrid network are specified in S3:

[0026] S3.1: Design of the TimeMixer model.

[0027] The TimeMixer model employs a full MLP architecture, achieving state-of-the-art performance in short-term prediction tasks by decoupling multi-scale time series information. Its core idea is to abandon the complex Transformer structure and revert to MLP, capturing complex patterns in sequences through innovative time-dimensional mixing and frequency domain decomposition. The model first inputs historical sequences and known future sequences, performing instance normalization to eliminate dimensional differences between sequences and enhance generalization ability. The TimeMixer model uses multi-scale downsampling, applying average pooling to downsample the input sequence at m levels, generating m-scale sequences. ,satisfy

[0028] (6)

[0029] Where P is the input length and C is the number of variables; m typically takes values ​​of 1–3, with 3 for long sequences and 1 for short sequences, balancing performance and efficiency. Then, a linear layer maps the sequences at each scale to a high-dimensional feature space, superimposed with sinusoidal positional encoding to preserve temporal information, outputting a multi-scale feature matrix. The core of the TimeMixer model consists of two symmetric but functionally different MLP modules, processing past and future information respectively: the PDM and FMM modules. The PDM module extracts historical information by mixing the decomposed seasonal and trend components at different scales. Since seasonal and trend components exhibit significantly different characteristics in time series, and time series at different scales reflect different attributes, it is necessary to decouple these two components at different scales. For each scale feature... The trend term is extracted using a large window moving average, and the seasonal term is the residual between the original feature and the trend term. The formula is as follows:

[0030] (7)

[0031] (8)

[0032] For trend items, It is a sliding window. This represents the seasonal component. Next, a two-way scale mixing method is used, with the seasonal component aggregated from the bottom up, gradually converging high-frequency information from the fine scale to the coarse scale, as shown in the following formula:

[0033] (9)

[0034] Downsampling and feature fusion are implemented using 3×3 convolutions, with a stride of s=2 to match the scale factor, and residuals are added to maintain gradient stability. Calculations are performed sequentially from m=0 to m=M−1, ultimately yielding the fused seasonal set. The trend components are applied from top to bottom, with the coarse-scale global trend guiding the fine-scale local modeling, as shown in the following formula:

[0035] (10)

[0036] Upsampling is achieved using a 2×2 transposed convolution, with a stride of s=2 to restore the scale, and residuals are added to ensure information consistency. The calculation proceeds sequentially from m=M to m=1, ultimately yielding the fused trend set. The combined seasonal and trend terms are recombined and then processed through a feedforward neural network to enhance channel interaction, as shown in the following formula.

[0037] (11)

[0038] (12)

[0039] (13)

[0040] It represents the fusion characteristics after recombination, restoring the complete temporal information at this scale. This represents the output features of the feedforward neural network, achieving nonlinear transformation along the channel dimension. Finally, residual connections and layer normalization are performed. The FMM block, based on the multi-scale features output by the PDM, generates the final prediction through weighted fusion of multiple predictors, balancing the periodic contributions of different scales. Features at each scale... Configure an independent linear predictor to map features to a prediction length F, as shown in the following formula:

[0041] (14)

[0042] This represents the prediction result at the m-th scale. The linear predictor parameters are learned independently to adapt to the periodic characteristics of each scale, and then a multi-scale prediction set is generated. Next, weighted fusion is performed, generating scale weights through a small network, as shown in the following formula.

[0043] (15)

[0044] Global average pooling is used to compress features, MLP outputs weights, and Softmax normalization ensures that the sum of the weights is 1. Finally, the fused predictions are... Mapping back to the original data scale, the formula is as follows:

[0045] (16)

[0046] μ and σ represent the mean and standard deviation of the preprocessing stage. The core of the TimeMixer model is composed of stacked PDM and FFM blocks. The core architecture of each modal component derived from S2 remains unchanged, while parameters are adjusted specifically for different frequency modes. The configurations for high-frequency, mid-frequency, and low-frequency modal parameters are as follows:

[0047]

[0048]

[0049]

[0050] Establish predictive sub-models that match different modal components.

[0051] S3.2: Reconstruction Method for Multimodal Components

[0052] This invention employs weighted linear reconstruction, which assigns weights based on the energy proportion of each mode in the original sequence. Modes with higher energy have larger weights, as shown in the following formula:

[0053] (17)

[0054] Weight calculation:

[0055] (18)

[0056] (19)

[0057] Let N represent the energy (mean square magnitude) of the i-th mode, and N be the sequence length. This represents the TimeMixer prediction value for the i-th mode. The advantage of this reconstruction method is its clear physical meaning; when the energy proportion of high-frequency or mid-frequency modes is high in wind power scenarios, the weights naturally match the prediction priority.

[0058] Traditional Transformer architectures are complex and inefficient when processing wind power sequences. In contrast, the TimeMixer model, lacking a self-attention mechanism, is lightweight and efficient, making it suitable for rapid modeling of wind power data. The TimeMixer model employs multi-scale decomposition and bidirectional mixing, simultaneously capturing hourly fluctuations and daily periodic variations in wind power. Furthermore, the TimeMixer model is robust, with residual connections and layer normalization reducing training difficulty and exhibiting good adaptability to noise and outliers in wind power data.

[0059] This completes the design based on adaptive mode decomposition and multi-scale hybrid time network, enabling short-term wind power prediction modeling.

[0060] The short-term wind power prediction model established in step S3 lays the foundation for subsequent engineering applications. In step S4, the feasibility of the established model is verified and analyzed using a simulation platform. The specific process is as follows:

[0061] S4.1: Based on one year of historical actual operation data of a wind farm unit in my country with hourly resolution, the model for short-term wind power prediction designed in S3 based on adaptive mode decomposition and multi-scale hybrid time network was selected and its effectiveness was verified.

[0062] S4.2: Based on the model established in step S3 and the ideas proposed in S4.1, plot the fitting curve and error curve between the actual wind power output and the established model.

[0063] S4.3: After model identification is completed through S3, calculate the mean absolute error (MAE), root mean square error (RMSE), and fitting accuracy of the model output. ), calculate the accuracy of the model.

[0064] (18)

[0065] (19)

[0066] (20)

[0067] S4.4: To highlight the accuracy and training speed of the adaptive mode decomposition and multi-scale hybrid time network model, the constructed model is compared with the traditional Transformer algorithm and the TimeMixer model without adaptive mode decomposition to verify the feasibility of the designed modeling method.

[0068] S4.5: To verify the robustness of the established model, noise was added to the wind power data before training the model and plotting the results. Beneficial effects of this invention:

[0069] This invention adaptively optimizes key parameters involved in the modal decomposition process, avoiding the instability of decomposition results caused by relying on manual experience to set the number of decomposition layers and penalty factors in existing technologies. By performing adaptive modal decomposition on the original wind power sequence, it breaks it down into several power subsequences with distinct frequency characteristics, transforming the original complex non-stationary signal into multiple relatively stationary modal components. This processing method effectively reduces noise interference and multi-scale feature aliasing in the sequence, making each modal component clearer in terms of temporal structure and statistical characteristics, which is beneficial to improving the learning effect of the prediction model on the variation law of wind power.

[0070] This invention does not employ a uniform prediction model structure for all modal components. Instead, it constructs a multi-scale temporal feature hybrid prediction sub-model that matches the frequency characteristics and time scale differences of each modal component. This approach enables the prediction model to effectively model both low-frequency trend modes and high-frequency fluctuation modes separately, avoiding the underfitting or overfitting problems caused by the single model structure in existing technologies, thus improving prediction performance from a structural perspective.

[0071] This invention effectively avoids the problem of gradual error accumulation in long-term series predictions by predicting each modal component separately and reconstructing the prediction results. The modal-level prediction and reconstruction mechanism can distribute prediction errors across different time scales for processing, thereby reducing prediction bias as a whole and improving the accuracy and stability of short-term wind power prediction results. Attached Figure Description

[0072] Figure 1 This is a flowchart of the slime mold optimization algorithm used in this invention to find the optimal number of modes and penalty factor in variational mode decomposition;

[0073] Figure 1 The text details the process of the slime mold optimization algorithm in finding optimal parameters. First, the core parameters of the algorithm are set, and then the initial positions of individual slime molds are initialized. Next, the fitness of each slime mold individual is calculated, and all individuals are sorted. Then, the algorithm's weights and key parameters are updated, and the slime mold positions are updated again. Finally, it checks if the stopping condition is met. If not, it returns to the fitness calculation and sorting steps to continue iterating; if the condition is met, it outputs the current optimal solution, i.e., the slime mold position with the best fitness.

[0074] Figure 2This is a simplified structural diagram of the wind power data sequence decomposition involved in this invention;

[0075] Figure 2 This includes adaptive mode decomposition to separate modal components at different scales. The raw wind power data is decomposed by SMA-VMD to obtain components (IMFs) with different center frequencies. The dominant frequency distribution and sample entropy are calculated for each IMF. Based on the dominant frequency distribution and sample entropy, the components are divided into low, medium and high frequency components.

[0076] Figure 3 This is a simplified structural diagram of TimeMixer, the short-term wind power prediction model involved in this invention.

[0077] Figure 3 The main components include the core modules of the TimeMixer model, PDM and FMM. The model first downsamples the wind power sequence to obtain a multi-scale sequence, then decomposes it to obtain a seasonal mixed sequence and a trend mixed sequence. Then, it is input into the predictor through a forward feedback network, and finally summed to obtain the predicted value.

[0078] Figure 4 This is a block diagram of the structure of the adaptive mode decomposition and multi-scale hybrid time network involved in this invention;

[0079] Figure 4 The original wind power data is first subjected to SMA-VMD adaptive decomposition, and then TimeMixer sub-models are established for low, medium and high frequency components respectively. Wind speed and wind direction are selected as features to be input into the prediction model. Finally, the prediction values ​​obtained from the three sub-models are weighted and fused to obtain the final prediction value. Detailed Implementation

[0080] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0081] Please refer to the accompanying drawings in the instruction manual. Figure 1 , Figure 1 This is a simplified flowchart of the slime mold optimization algorithm involved in this invention. The process begins with "Start," where core parameters such as population size are set, and initial positions of slime mold individuals are initialized within the search space to form the initial population. Next, the fitness of each slime mold individual is calculated, and the merits of each position are evaluated based on the optimization objective, and all individuals are sorted. Then, the algorithm's weights and key parameters are updated based on the current fitness, and the slime molds are moved to better regions based on the new parameters to update their positions. Finally, it is determined whether the stopping condition is met. If not, the iteration returns to the fitness calculation and sorting steps; if the condition is met, the current optimal solution, i.e., the slime mold position with the best fitness, is output, thus determining the optimal k and α for VMD decomposition.

[0082] Please refer to the accompanying drawings in the instruction manual.Figure 2 , Figure 2 This is a simplified structural diagram of the wind power data sequence decomposition involved in this invention. First, the original wind power data is input. The number of parameter modes and penalty factor are determined using the SMA algorithm. Then, these parameters are substituted into VMD decomposition to decompose the original data into multiple IMF components, such as IMF1, IMF2, IMF3, etc. Next, based on the dominant frequency distribution and sample entropy of each IMF component, they are classified and integrated into three types of mode components: low-frequency, mid-frequency, and high-frequency. This achieves hierarchical decomposition and classification of the wind power data, preparing for the subsequent establishment of prediction sub-models.

[0083] Please refer to the accompanying drawings in the instruction manual. Figure 3 , Figure 3 This is a simplified structural diagram of the TimeMixer short-term wind power prediction model involved in this invention. First, there is a multi-scale time series module, which downsamples the input modal component sequence into subsequences of different scales to prepare for subsequent decomposition. Next is the historical decomposable mixing module, which first decomposes the multi-scale input sequence into subsequences of corresponding scales, and then processes them along two paths. Seasonal mixing integrates fine-grained to coarse-grained seasonal fluctuation features in a bottom-up manner, while trend mixing transmits coarse-grained to fine-grained trend features in a top-down manner. Finally, the two types of features are merged, and the integrated historical features are output through a feedforward module. Finally, there is a future multi-predictor mixing module, which inputs the feedforward features into multiple predictors, each predictor corresponding to an output sequence at one scale. The prediction results from all scales are then merged to obtain the final prediction result.

[0084] Please refer to the accompanying drawings in the instruction manual. Figure 4 , Figure 4 This is a block diagram of the structure based on adaptive mode decomposition and multi-scale hybrid time network mentioned in this invention. Based on a wind farm turbine in my country, the method steps include:

[0085] S1: Adaptively search for key parameters of variational mode decomposition using an intelligent optimization algorithm. The key parameters include the number of components in the mode decomposition and the penalty factor.

[0086] S2: Perform multimodal decomposition on the wind power sequence to obtain several power subsequences with different frequency characteristics. Based on the dominant frequency distribution of the modal components and the sample entropy, divide the power subsequences into modal components of different time scales.

[0087] S3: For different modal components, construct a multi-scale time feature hybrid prediction sub-model that matches their time scale characteristics. Finally, reconstruct the prediction results of each modal component by weighting to obtain the final prediction result of wind power.

[0088] S4: Verify and analyze the feasibility of the proposed prediction method using a simulation platform.

[0089] S1: The performance of Variational Mode Decomposition (VMD) heavily depends on the number of modes K and the penalty factor α. Determining these parameters through experience and trial and error increases workload and does not guarantee optimal results. The slime mold optimization algorithm is a metaheuristic optimization algorithm inspired by the foraging behavior of slime molds in nature. It completely avoids manual parameter selection and achieves adaptive and intelligent selection of VMD parameters. Slime molds form interconnected path networks through oscillations, achieving a balance between exploring food sources (global search) and narrowing paths to concentrate resources on high-quality food (local development). VMD parameter selection is defined as an optimization problem, and the VMD parameters K and α to be optimized are encoded as position vectors of individual slime molds in the optimization algorithm. First, the position vectors of individual slime molds are randomly initialized in the search space. Then, the objective function value, i.e., the fitness, corresponding to each individual's position is calculated. The fitness function design is the most crucial step; its goal is to minimize an index that measures the quality of decomposition. This invention uses envelope entropy as the index, which reflects the sparsity and impact characteristics of the signal. The smaller the entropy, the purer the decomposed modes. The weights W and oscillation parameters p are updated to simulate the propagation waves generated by the biological oscillator during slime mold searching. W dynamically changes based on the number of iterations and the individual's fitness ranking. Individuals with higher fitness have higher weights, attracting other individuals to gravitate towards them. During the global search phase, when the random number is greater than the parameter p, the individual's position is less influenced by a randomly selected individual or the current best individual, resulting in more random searching. During the local search phase, when the random number is less than or equal to p, individuals strongly move towards the current best individual, performing a refined local search. The position update formulas for global and local searches are as follows:

[0090] (1)

[0091] (2)

[0092] Indicates a random position. , This represents two distinct random individuals used to construct the search direction. This represents the individual's previous position, and vb and vc are oscillation parameters. The fitness of the new position is compared with that of the old position, the better position is retained, and the iteration is repeated until the stopping condition is met.

[0093] S2: Based on S1, when the slime mold optimization algorithm reaches the stopping condition, the globally optimal individual, i.e., the optimal VMD parameter combination, is output. Using this set of optimal parameters, the final VMD decomposition is performed on the original signal to obtain the optimal modal components. VMD is a completely non-recursive signal decomposition method. Its core idea is to adaptively decompose the original signal into K quasi-orthogonal modal functions with specific sparsity. Its core objective is to find a set of modes. and the corresponding center frequency This minimizes the sum of bandwidths for all modes while ensuring accurate reconstruction of the original signal. Modal decomposition is performed on the wind power sequence to obtain subsequences of different frequencies. Then, based on the dominant frequency distribution and sample entropy, the power subsequences are divided into modal components at different time scales. The specific steps are as follows:

[0094] S2.1: First, initialize all modes. Center frequency Lagrange multipliers The value is 0. In the frequency domain, all modes are updated. The update formula for the k-th mode is:

[0095] (3)

[0096] The molecular part is derived from the original signal spectrum. In, subtract all other modes The current estimate, plus the Lagrange multipliers. The denominator is a fraction... The filter is centered at the center frequency. It acts as a bandpass filter; near the center frequency, the denominator is small, and the overall fractional value is large, allowing that frequency component to pass. Far from the center frequency, the denominator increases sharply, the fractional value decreases, and that frequency component is suppressed. The penalty factor α controls the filter's bandwidth; the larger α is, the faster the denominator grows with frequency shift, the narrower the filter, and the smaller the modal bandwidth. Then, all center frequencies are updated. The center frequency is updated by calculating the first moment of the power spectrum of that mode:

[0097] (4)

[0098] Energy distribution of modes Viewed as a density function, the new center frequency is the location of the first moment of this density function. This allows the center frequency of each mode to automatically move to the frequency band where its energy is most concentrated. Furthermore, the Lagrange multipliers need to be updated to strengthen the reconstruction constraints, as shown in the following formula:

[0099] (5)

[0100] This is the update step size, typically taken as the noise tolerance parameter, defaulting to 0 to ensure strict constraint satisfaction. The term in parentheses is the spectrum of the reconstructed residual in the current iteration. Finally, the convergence condition is checked, and k time-domain modal components and their corresponding center frequencies are output.

[0101] S2.2: Following S2.1, k modal components and their corresponding center frequencies are obtained. High frequencies correspond to short-term rapid fluctuations, while low frequencies correspond to long-term slow trends. In addition to the center frequencies, this invention also uses sample entropy to quantify the complexity and randomness of each modal component. Given the embedding dimension m and tolerance r, sample entropy statistically calculates the probability of similarity between vector templates of length m and m+1 in the time series. The larger the sample entropy value, the more complex, irregular, and unpredictable the sequence. Therefore, modal components with high sample entropy can be considered noise, while modal components with low sample entropy are retained. Adaptive modal decomposition separates the originally complex wind power sequence into subsequences with different scale characteristics. Then, based on the dominant frequency distribution and sample entropy, the subsequences are classified to finally obtain effective high-frequency, mid-frequency, and low-frequency modal components.

[0102] After obtaining modal components from the wind power sequence through adaptive mode decomposition, the design steps of the multi-scale time hybrid network are specified in S3:

[0103] S3.1: Design of the TimeMixer model.

[0104] The TimeMixer model employs a full MLP architecture, achieving state-of-the-art performance in short-term prediction tasks by decoupling multi-scale time series information. Its core idea is to abandon the complex Transformer structure and revert to MLP, capturing complex patterns in sequences through innovative time-dimensional mixing and frequency domain decomposition. The model first inputs historical sequences and known future sequences, performing instance normalization to eliminate dimensional differences between sequences and enhance generalization ability. The TimeMixer model uses multi-scale downsampling, applying average pooling to downsample the input sequence at m levels, generating m-scale sequences. ,satisfy

[0105] (6)

[0106] Where P is the input length and C is the number of variables; m typically takes values ​​of 1–3, with 3 for long sequences and 1 for short sequences, balancing performance and efficiency. Then, a linear layer maps the sequences at each scale to a high-dimensional feature space, superimposed with sinusoidal positional encoding to preserve temporal information, outputting a multi-scale feature matrix. The core of the TimeMixer model consists of two symmetric but functionally different MLP modules, processing past and future information respectively: the PDM and FMM modules. The PDM module extracts historical information by mixing the decomposed seasonal and trend components at different scales. Since seasonal and trend components exhibit significantly different characteristics in time series, and time series at different scales reflect different attributes, it is necessary to decouple these two components at different scales. For each scale feature... The trend term is extracted using a large window moving average, and the seasonal term is the residual between the original feature and the trend term. The formula is as follows:

[0107] (7)

[0108] (8)

[0109] For trend items, It is a sliding window. This represents the seasonal component. Next, a two-way scale mixing method is used, with the seasonal component aggregated from the bottom up, gradually converging high-frequency information from the fine scale to the coarse scale, as shown in the following formula:

[0110] (9)

[0111] Downsampling and feature fusion are implemented using 3×3 convolutions, with a stride of s=2 to match the scale factor, and residuals are added to maintain gradient stability. Calculations are performed sequentially from m=0 to m=M−1, ultimately yielding the fused seasonal set. The trend components are applied from top to bottom, with the coarse-scale global trend guiding the fine-scale local modeling, as shown in the following formula:

[0112] (10)

[0113] Upsampling is achieved using a 2×2 transposed convolution, with a stride of s=2 to restore the scale, and residuals are added to ensure information consistency. The calculation proceeds sequentially from m=M to m=1, ultimately yielding the fused trend set. The combined seasonal and trend terms are recombined and then processed through a feedforward neural network to enhance channel interaction, as shown in the following formula.

[0114] (11)

[0115] (12)

[0116] (13)

[0117] It represents the fusion characteristics after recombination, restoring the complete temporal information at this scale. This represents the output features of the feedforward neural network, achieving nonlinear transformation along the channel dimension. Finally, residual connections and layer normalization are performed. The FMM block, based on the multi-scale features output by the PDM, generates the final prediction through weighted fusion of multiple predictors, balancing the periodic contributions of different scales. Features at each scale... Configure an independent linear predictor to map features to a prediction length F, as shown in the following formula:

[0118] (14)

[0119] This represents the prediction result at the m-th scale. The linear predictor parameters are learned independently to adapt to the periodic characteristics of each scale, and then a multi-scale prediction set is generated. Next, weighted fusion is performed, generating scale weights through a small network, as shown in the following formula.

[0120] (15)

[0121] Global average pooling is used to compress features, MLP outputs weights, and Softmax normalization ensures that the sum of the weights is 1. Finally, the fused predictions are... Mapping back to the original data scale, the formula is as follows:

[0122] (16)

[0123] μ and σ represent the mean and standard deviation of the preprocessing stage. The core of the TimeMixer model is composed of stacked PDM and FFM blocks. The core architecture of each modal component derived from S2 remains unchanged, while parameters are adjusted specifically for different frequency modes. The configurations for high-frequency, mid-frequency, and low-frequency modal parameters are as follows:

[0124]

[0125]

[0126]

[0127] Establish predictive sub-models that match different modal components.

[0128] S3.2: Reconstruction Method for Multimodal Components

[0129] This invention employs weighted linear reconstruction, which assigns weights based on the energy proportion of each mode in the original sequence. Modes with higher energy have larger weights, as shown in the following formula:

[0130] (17)

[0131] Weight calculation:

[0132] (18)

[0133] (19)

[0134] Let N represent the energy (mean square magnitude) of the i-th mode, and N be the sequence length. This represents the TimeMixer prediction value for the i-th mode. The advantage of this reconstruction method is its clear physical meaning; when the energy proportion of high-frequency or mid-frequency modes is high in wind power scenarios, the weights naturally match the prediction priority.

[0135] Traditional Transformer architectures are complex and inefficient when processing wind power sequences. In contrast, the TimeMixer model, lacking a self-attention mechanism, is lightweight and efficient, making it suitable for rapid modeling of wind power data. The TimeMixer model employs multi-scale decomposition and bidirectional mixing, simultaneously capturing hourly fluctuations and daily periodic variations in wind power. Furthermore, the TimeMixer model is robust, with residual connections and layer normalization reducing training difficulty and exhibiting good adaptability to noise and outliers in wind power data.

[0136] This completes the design based on adaptive mode decomposition and multi-scale hybrid time network, enabling short-term wind power prediction modeling.

[0137] The short-term wind power prediction model established in step S3 lays the foundation for subsequent engineering applications. In step S4, the feasibility of the established model is verified and analyzed using a simulation platform. The specific process is as follows:

[0138] S4.1: Based on one year of historical actual operation data of a wind farm unit in my country with hourly resolution, the model for short-term wind power prediction designed in S3 based on adaptive mode decomposition and multi-scale hybrid time network was selected and its effectiveness was verified.

[0139] S4.2: Based on the model established in step S3 and the ideas proposed in S4.1, plot the fitting curve and error curve between the actual wind power output and the established model.

[0140] S4.3: After model identification is completed through S3, calculate the mean absolute error (MAE), root mean square error (RMSE), and fitting accuracy of the model output. ), calculate the accuracy of the model.

[0141] (18)

[0142] (19)

[0143] (20)

[0144] S4.4: To highlight the accuracy and training speed of the adaptive mode decomposition and multi-scale hybrid time network model, the constructed model is compared with the traditional Transformer algorithm and the TimeMixer model without adaptive mode decomposition to verify the feasibility of the designed modeling method.

[0145] S4.5: To verify the robustness of the established model, noise was added to the wind power data before training the model and plotting the results.

Claims

1. A short-term wind power prediction method based on adaptive mode decomposition and multi-scale time hybrid network, characterized in that: Includes the following steps: S1: Adaptively search for key parameters of variational mode decomposition using an intelligent optimization algorithm. The key parameters include the number of components in the mode decomposition and the penalty factor. S2: Perform multimodal decomposition on the wind power sequence to obtain several power subsequences with different frequency characteristics. Based on the dominant frequency distribution of the modal components and the sample entropy, divide the power subsequences into modal components of different time scales. S3: For different modal components, construct a multi-scale time feature hybrid prediction sub-model that matches their time scale characteristics. Finally, reconstruct the prediction results of each modal component by weighting to obtain the final prediction result of wind power. S4: Verify and analyze the feasibility of the proposed prediction method using a simulation platform.

2. The short-term wind power prediction method based on adaptive mode decomposition and multi-scale time hybrid network according to claim 1, characterized in that: The key challenges of adaptively searching for variational mode decomposition parameters using intelligent optimization algorithms in step S1 are as follows: S1: The performance of Variational Mode Decomposition (VMD) heavily depends on the number of modes K and the penalty factor α. Determining the optimal parameters through experience and trial and error increases workload and does not guarantee finding the best parameters. The slime mold optimization algorithm is a metaheuristic optimization algorithm inspired by the foraging behavior of slime molds in nature. It completely avoids manual parameter selection and achieves adaptive and intelligent selection of VMD parameters. VMD parameter selection is defined as an optimization problem, and the VMD parameters K and α to be optimized are encoded as position vectors of individual slime molds in the algorithm. First, the core parameters of the algorithm are set. Then, the position vectors of individual slime molds are randomly initialized in the search space. Next, the objective function value corresponding to each individual's position, i.e., the fitness, is calculated. The fitness function design is the most crucial step; its goal is to minimize an index that measures the quality of decomposition. This invention uses envelope entropy as an index, which reflects the sparsity and impact characteristics of the signal. The smaller the entropy, the purer the decomposed modes. The weights W and key parameters are updated. The formula for W is as follows: (1) Where SmellOrder is the individual fitness ranking index; then, the population position is updated using the following formula: (2) p is the probability threshold. Indicates a random position. Find the current optimal position; compare the fitness of the new position and the old position, retain the better one, and repeat the iteration until the stopping condition is met.

3. The short-term wind power prediction method based on adaptive mode decomposition and multi-scale time hybrid network according to claim 1, characterized in that: If the wind power sequence is decomposed into multiple modes and the power subsequence is divided into modal components of different time scales based on the dominant frequency distribution and sample entropy of the modal components, then step S2 can be specified as follows: S2.1: The original nonlinear and non-stationary complex wind power signal is adaptively decomposed into several intrinsic mode function (IMF) components with specific sparsity and non-overlapping characteristics. Each component corresponds to a center frequency and has approximately stationary characteristics. At the same time, the bandwidth and fidelity of the components are balanced by parameters such as mode number and penalty factor, effectively avoiding problems such as mode mixing and endpoint effects. S2.2: In S2.1, we obtain k modal components and their corresponding center frequencies. High frequencies correspond to short-term rapid fluctuations, while low frequencies correspond to long-term slow trends. In addition to the center frequency, this invention also uses sample entropy to quantify the complexity and randomness of each modal component; the larger the sample entropy value, the more complex, irregular and unpredictable the sequence; therefore, modal components with high sample entropy can be regarded as noise, while modal components with low sample entropy are retained; the original complex wind power sequence is separated into subsequences with different scale characteristics according to adaptive mode decomposition, and then the subsequences are classified according to the dominant frequency distribution and sample entropy to finally obtain effective high-frequency modal components, mid-frequency modal components and low-frequency modal components.

4. The short-term wind power prediction method based on adaptive mode decomposition and multi-scale time hybrid network according to claim 1, characterized in that: This paper describes the principle and design steps of a short-term wind power prediction method based on a multi-scale hybrid network: S3.1: Design of the TimeMixer model; The TimeMixer model employs a full MLP architecture, achieving state-of-the-art performance in short-term prediction tasks by decoupling multi-scale time series information. Its core idea is to abandon the complex Transformer structure and revert to MLP, capturing complex patterns in sequences through innovative time-dimensional mixing and frequency domain decomposition. The model first inputs historical sequences and known future sequences, performing instance normalization to eliminate dimensional differences between different sequences and enhance the model's generalization ability. The TimeMixer model uses multi-scale downsampling generation, applying average pooling to downsample the input sequence at m levels, generating m-scale sequences that satisfy… (3) Where P is the input length and C is the number of variables; m is usually 1–3, with 3 for long sequences and 1 for short sequences, balancing performance and efficiency; then, a linear layer maps the sequences at each scale to a high-dimensional feature space, and sinusoidal positional encoding is superimposed to preserve temporal information, outputting a multi-scale feature matrix; the core of the TimeMixer model consists of two symmetric but functionally different MLP modules, processing past and future information respectively, namely the PDM and FMM modules; the PDM module extracts historical information by mixing the seasonal and trend components at different scales; since seasonal and trend components exhibit significantly different characteristics in time series, and time series at different scales reflect different attributes, it is necessary to decouple these two components at different scales; for each scale feature, a large window moving average is used to extract the trend term, and the seasonal term is the residual between the original feature and the trend term, as shown in the following formula: (4) (5) For trend items, It is a sliding window. For the seasonal component, a two-way scale mixing method is then used, where the seasonal component is aggregated from the bottom up, with high-frequency information from the fine scale gradually aggregating to the coarse scale, as shown in the following formula: (6) Downsampling and feature fusion are implemented using 3×3 convolutions, with a stride of s=2 to match the scale factor, and residuals are added to maintain gradient stability; calculations are performed sequentially from m=0 to m=M−1, ultimately yielding the fused seasonal set. The trend components are analyzed from top to bottom, with the coarse-scale global trend guiding the fine-scale local modeling. The formula is as follows: (7) The mixed seasonal and trend terms are recombined, and channel interactions are enhanced via a feedforward neural network. Finally, residual connections and layer normalization are performed. The FMM block generates the final prediction based on the multi-scale features output by the PDM through weighted fusion of multiple predictors, balancing the periodic contributions of different scales. An independent linear predictor is configured for the features at each scale, mapping the features to the prediction length F, as shown in the following formula: (8) The prediction result at scale m is represented by the linear predictor parameters, which are learned independently to adapt to the periodic characteristics of each scale, and then a multi-scale prediction set is generated. Next, weighted fusion is performed, with scale weights generated through a small network, as shown in the following formula: (9) Global average pooling is used to compress features, MLP outputs weights, and Softmax normalization ensures that the sum of the weights is 1; finally, the fused predictions are mapped back to the original data scale, as shown in the following formula: (10) μ and σ represent the mean and standard deviation of the preprocessing stage; the core of the TimeMixer model is composed of stacked PDM and FFM blocks; the core architecture of each modal component decomposed by S2 remains unchanged, and the parameters are adjusted specifically for different frequency modes; the configurations of high-frequency modal parameters, mid-frequency modal parameters, and low-frequency modal parameters are as follows: Establish predictive sub-models that match different modal components; S3.2: Reconstruction method for multimodal components; This invention employs weighted linear reconstruction, which assigns weights based on the energy proportion of each mode in the original sequence. Modes with higher energy have larger weights, as shown in the following formula: (11) Weight calculation: (12) (13) Represents the energy (mean square amplitude) of the i-th mode, where N is the sequence length; represents the TimeMixer prediction value of the i-th mode; the advantage of this reconstruction method is that the physical meaning is clear, and when the energy of high-frequency or mid-frequency modes accounts for a high proportion in wind power scenarios, the weights naturally match the prediction priority; Traditional Transformer architectures are complex and inefficient when processing wind power sequences. In contrast, the TimeMixer model, which lacks a self-attention mechanism, is lightweight and efficient, making it suitable for rapid modeling of wind power data. The TimeMixer model employs multi-scale decomposition and bidirectional mixing, enabling it to simultaneously capture hourly fluctuations and daily periodic variations in wind power. Furthermore, the TimeMixer model is robust, with residual connections and layer normalization design reducing training difficulty and exhibiting good adaptability to noise and outliers in wind power data. This completes the design based on adaptive mode decomposition and multi-scale hybrid time network, enabling short-term wind power prediction modeling.

5. The short-term wind power prediction method based on adaptive mode decomposition and multi-scale time hybrid network according to claim 1, characterized in that: The feasibility of the proposed modeling method is verified and analyzed using a simulation platform. In step S4, the feasibility of the established model is verified and analyzed using a simulation platform. The specific process is as follows: S4.1: Based on one year of historical actual operation data of a wind farm unit in my country with hourly resolution, the model for short-term wind power prediction designed in S3 based on adaptive mode decomposition and multi-scale hybrid time network was selected and its effectiveness was verified. S4.2: Based on the model established in step S3 and the ideas proposed in S4.1, plot the fitting curve and error curve between the actual wind power output and the established model; S4.3: After model identification is completed through S3, calculate the mean absolute error (MAE), root mean square error (RMSE), and fitting accuracy of the model output. ), calculate the accuracy of the model; (14) (15) (16) S4.4: To highlight the accuracy and training speed of the adaptive mode decomposition and multi-scale hybrid time network model, the model is compared with the traditional Transformer algorithm and the TimeMixer model without adaptive mode decomposition to verify the feasibility of the designed modeling method. S4.5: To verify the robustness of the established model, noise was added to the wind power data before training the model and plotting the results.

Images