A wind power prediction method and system based on time-frequency cooperation and adaptive decoupling

Abstract

This invention relates to the field of new energy power prediction technology in power systems, specifically a wind power prediction method and system based on time-frequency coordination and adaptive decoupling. The method includes: acquiring historical wind power data; performing complexity-aware adaptive modal decomposition and reconstruction on the data to obtain sets of reconstructed sub-signals of different complexities; inputting each sub-signal set into a heterogeneous feature extraction network for parallel feature extraction to obtain heterogeneous features; adaptively weighting and fusing the heterogeneous features through a dynamic gated fusion network to generate fused features and obtain a preliminary prediction; inputting the preliminary prediction result into a frequency domain closed-loop correction network, correcting its spectrum through a learnable frequency domain filter, and converting it back to the time domain for residual compensation to obtain the final prediction. This invention achieves accurate decoupling, intelligent fusion, and in-depth correction of wind power sequences, thereby improving the accuracy of wind power prediction.

Description

Technical Field

[0001] This invention relates to the field of new energy power prediction technology in power systems, specifically a wind power prediction method and system based on time-frequency coordination and adaptive decoupling. Background Technology

[0002] In the context of building a new power system, high-precision wind power forecasting is crucial for ensuring the safe and stable operation of the power grid. However, wind power sequences are affected by meteorological factors, exhibiting nonlinearity, non-stationarity, and multi-scale characteristics. Currently, forecasting methods based on a hybrid architecture of "decomposition-integration" and deep learning are the main technical approach, but existing technologies still have the following shortcomings:

[0003] 1. Homogenization problem in modal decoupling: Existing techniques such as variational mode decomposition (VMD) rely heavily on empirical parameter settings, which can easily lead to mode aliasing. Furthermore, all decomposed sub-components are homogenized and input into the same prediction model, ignoring the differences in dynamic properties between high-frequency noise and low-frequency trends, resulting in low feature extraction efficiency.

[0004] 2. Static adaptation problem in feature fusion: Existing technologies often use a combination of Transformer and RNN architectures to capture long-range and short-range dependencies, but feature fusion often uses simple concatenation or fixed weighting, which cannot perceive dynamic changes in the sequence. This can easily introduce noise during periods of stable wind conditions and lead to response lag during periods of sudden changes.

[0005] 3. Open-loop problem of error correction: Error accumulation exists in multi-step prediction. Existing error correction methods are mostly treated as independent post-processing modules, which are separated from the feature extraction process of the backbone prediction model, making it difficult for the corrector to suppress error propagation mechanistically.

[0006] Therefore, how to provide a wind power prediction method that can overcome the above-mentioned defects has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a wind power prediction method and system based on time-frequency coordination and adaptive decoupling, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A wind power prediction method based on time-frequency coordination and adaptive decoupling includes the following steps:

[0010] S1. Obtain historical wind power data;

[0011] S2. Perform complexity-aware adaptive mode decomposition and recombination on the historical wind power data to obtain multiple sets of reconstructed sub-signals with different dynamic complexity characteristics;

[0012] S3. Input the multiple reconstructed sub-signal sets into a heterogeneous feature extraction network for parallel feature extraction to obtain heterogeneous features;

[0013] S4. Adaptively weighted fusion of the heterogeneous features is performed through a dynamic gated fusion network to generate fused features, and preliminary power prediction results are obtained based on the fused features.

[0014] S5. Input the preliminary power prediction result into the frequency domain closed-loop correction network, correct the spectrum of the preliminary power prediction result through a learnable frequency domain filter, convert the corrected spectrum back to the time domain, and compensate the preliminary power prediction result in the form of residuals to obtain the final power prediction result.

[0015] In particular, steps S2, S3, S4 and S5 are performed through collaborative training using an end-to-end joint loss function.

[0016] As a further aspect of the present invention, step S2 specifically includes:

[0017] S21. With minimizing the average envelope entropy of each component after decomposition as the optimization objective, adaptively determine the number of modes K and the penalty factor α of variational mode decomposition (VMD);

[0018] S22. Using the determined number of modes K and the penalty factor α, variational mode decomposition is performed on the historical wind power data to obtain multiple intrinsic mode function components;

[0019] S23. Calculate the sample entropy of each intrinsic mode function component, and perform unsupervised clustering and recombination on all components based on the sample entropy to form the reconstructed sub-signal set, which includes at least a low-complexity subset and a high-complexity subset.

[0020] As a further aspect of the present invention: step S3 specifically includes:

[0021] S31. Input the low-complexity subset into the global feature extractor to obtain long-range dependency features;

[0022] S32. Input the high-complexity subset into the local feature extractor to obtain local transient features;

[0023] S33. Align the dimensions of the local transient features with those of the long-range dependent features.

[0024] As a further aspect of the present invention: step S4 specifically includes:

[0025] S41. The long-range dependency features are concatenated with the dimension-aligned local transient features to obtain joint context features;

[0026] S42. Input the joint context features into the gating network to generate a dynamic weight matrix G with the same dimension as the joint context features. The output value range of the gating network is (0, 1).

[0027] S43. Using the dynamic weight matrix, perform element-wise weighted summation on the long-range dependency feature and the local transient feature to obtain the fused feature F: , where H g H is a long-range dependency feature. l These are local transient features after dimension alignment. This represents element-wise multiplication. To be related to the dynamic weight matrix A matrix of all dimensions consisting entirely of 1s, meaning that all elements in the matrix have a value of 1.

[0028] As a further aspect of the present invention: step S5 specifically includes:

[0029] S51. Perform a real-number fast Fourier transform on the preliminary power prediction result and map it to the frequency domain to obtain the frequency domain prediction result.

[0030] S52. Multiply the frequency domain prediction result element-wise with the learnable complex filter to obtain the corrected spectrum, where each element of the complex filter... model and argument These are used to adjust the amplitude and phase of the corresponding frequency components, respectively. The amplitude weights are assigned to the corresponding frequency positions. Here, represents the phase offset at the corresponding frequency position; j is the imaginary unit, used to characterize the phase information of the complex exponential function, which is adjusted... and This enables joint correction of the amplitude and phase responses of the corresponding frequency components;

[0031] S53. Perform a real-number fast Fourier inverse transform on the corrected spectrum and take the real part to obtain the time-domain correction term;

[0032] S54. Add the time-domain correction term to the preliminary power prediction result to obtain the final power prediction result.

[0033] As a further aspect of the present invention: step S54 specifically comprises:

[0034]

[0035] in, For the final power prediction results, This is the preliminary power prediction result, where Δ is the time-domain correction term. A learnable scaling factor. This represents the actual, objective wind power output measured at the wind farm. This represents the result calculated and derived by an algorithm (neural network). It is an "estimate" or "prediction" made by the model based on historical data, not a definite fact directly read by physical sensors, so it must be labeled in mathematics to distinguish it.

[0036] As a further aspect of the present invention: the end-to-end joint loss function The final power prediction result is weighted and consists of a first loss between the initial power prediction result and the actual value, and a second loss between the final power prediction result and the actual value.

[0037]

[0038] Where Y is the true value, β is the weighting coefficient, and the value of β ranges from 0 to 1. For preliminary prediction of auxiliary losses, This represents the final predicted main loss.

[0039] This invention also provides a wind power prediction system based on time-frequency coordination and adaptive decoupling, comprising:

[0040] The data preprocessing module is used to acquire and standardize the historical wind power sequence;

[0041] A complexity-aware decoupling module is used to perform adaptive mode decomposition and complexity-based recombination on the standardized wind power sequence;

[0042] The dynamic feature fusion prediction module includes: a heterogeneous feature extraction submodule, used to extract features from the recombined sub-signal set; and a dynamic gated fusion submodule, used to adaptively fuse the extracted heterogeneous features and output preliminary power prediction results.

[0043] The frequency domain closed-loop correction module is used to filter and correct the error of the preliminary power prediction result in the frequency domain, and reconstruct the final power prediction result.

[0044] The collaborative training module is used to collaboratively optimize the parameters of the complexity-aware decoupling module, the dynamic feature fusion prediction module, and the frequency domain closed-loop correction module through an end-to-end joint loss function.

[0045] As a further aspect of the present invention: the complexity-aware decoupling module includes:

[0046] The parameter optimization unit is used to optimize VMD parameters with the goal of minimizing the average envelope entropy of the decomposed components.

[0047] The decomposition and quantization unit is used to perform VMD decomposition and calculate the sample entropy of each component;

[0048] Clustering and recombination units are used to cluster components based on sample entropy and recombine them into multiple reconstructed subsets, including at least low-complexity subsets and high-complexity subsets.

[0049] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any of the preceding embodiments.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] 1. Precise source decoupling: By adaptively optimizing VMD parameters with the goal of minimizing the average envelope entropy, the modality mixing problem caused by the empirical parameterization of traditional decomposition methods is solved; combined with sample entropy clustering and recombination, the decomposition components are divided into three categories of low, medium and high dynamic complexity, and input into different feature extraction networks respectively, overcoming the shortcomings of traditional methods in homogenizing all components.

[0052] 2. Intelligent Feature Fusion: Through a dynamic gating fusion network, context-aware adaptive fusion of global attention features and local recurrent features is achieved. When the sequence is stable, the gating weights tend to be 1, and the model relies more on the global trend; when the sequence changes drastically, the gating weights tend to be 0, and the model focuses on local transients. This input-dependent dynamic modulation mechanism effectively improves the model's adaptability to different operating conditions.

[0053] 3. In-depth error correction: The time-domain error correction problem is transformed into a frequency-domain complex filtering problem. A learnable complex filter simultaneously corrects both amplitude attenuation and phase shift in the prediction results. Unlike existing open-loop correction methods, the correction module of this invention is jointly trained with the backbone network, enabling it to learn the spectral evolution of errors at a mechanistic level, fundamentally suppressing error accumulation in multi-step predictions.

[0054] 4. Framework Collaborative Optimization: This invention integrates three major modules—adaptive decoupling, dynamic fusion, and frequency domain correction—into a unified end-to-end joint training framework, breaking the chain-like defects of traditional methods that involve "decomposition → fusion → correction." The joint loss function simultaneously constrains both the initial and final predictions, forcing the backbone network to generate error patterns that are easily corrected by frequency domain filters, thus achieving optimal global performance.

[0055] 5. Significantly Improved Prediction Accuracy: Experimental results show that the method of this invention significantly outperforms existing mainstream models in core evaluation metrics such as MSE, MAE, RMSE, MAPE, and R². Compared with the baseline model, MSE is reduced by more than 96%, MAE by more than 80%, and R² reaches 0.9563. In the power abrupt change range, the method of this invention can smoothly and accurately track the true trajectory; in long-term prediction, it effectively suppresses error amplification and phase drift problems. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the wind power prediction method provided in an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram comparing the effects of the wind power prediction method provided in this embodiment of the invention with traditional methods.

[0058] Figure 3 This is a schematic diagram of the reconstructed sub-signal obtained after adaptive decomposition of the original wind power sequence in an embodiment of the present invention.

[0059] Figure 4 This is a schematic diagram of the convergence curve of the optimization algorithm used in the embodiments of the present invention.

[0060] Figure 5 This is a schematic diagram of the dynamic fusion mechanism of the heterogeneous feature extraction network in an embodiment of the present invention.

[0061] Figure 6 This is a graph showing the model training convergence performance in an embodiment of the present invention.

[0062] Figure 7 This is a curve showing the convergence performance of the comparative model in an embodiment of the present invention. Detailed Implementation

[0063] The technical solution of this application will be further described in detail below with reference to specific embodiments.

[0064] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0065] Example 1: Multi-step advance prediction system based on real wind farm data

[0066] Please see Figure 1 This embodiment uses multi-step advance prediction based on actual wind farm data to illustrate the technical solution of the present invention in detail.

[0067] S1: Data Acquisition and Preprocessing

[0068] S11: Obtain the historical sequence of wind power of the wind farm and the corresponding meteorological covariate data, wherein the meteorological covariate includes at least one of wind speed, wind direction, temperature and air pressure.

[0069] S12: Perform Z-Score standardization on all data to ensure that the target variable (wind power) and multidimensional meteorological covariates use independent normalization operators to prevent data leakage.

[0070] S13: Construct autoregressive sample pairs based on sliding window technology, and strictly divide the training set and test set.

[0071] S2: Complexity-Aware Adaptive Modal Decomposition and Reorganization

[0072] S21: Adaptive parameter optimization

[0073] The White Shark Optimization (WSO) algorithm is set to a population size of 8 and a maximum number of iterations of 10. The fitness function is set to the minimum mean envelope entropy of the resolved signal envelopes of each IMF in the VMD solution. The VMD decomposition can be represented as a function... ,

[0074] in, Indicates the input wind power sequence; Represents the number of modes; This represents the penalty factor. Output Each intrinsic mode function (IMF). This invention models parameter selection as an optimization problem: finding the optimal combination of parameters (...). That is, searching for the optimal combination of parameters within the candidate parameter space. ).in, and These represent the optimal number of modes and the optimal penalty factor obtained after optimization, respectively; and are consistent with the above. and compared to, and The parameters to be searched. and The optimal parameters are determined. The fitness function is constructed to minimize the average envelope entropy of each IMF after decomposition.

[0075]

[0076] in, It is the envelope entropy of the k-th IMF. Its Hilbert envelope is in the 1st Normalized probability distribution at each sampling point Where N is the sample length of the wind power sequence, minimizing the average envelope entropy forces the resulting IMFs to be the "purest," with the lowest randomness and the best band-limited characteristics. This provides downstream prediction models with input features that are less noisy and more regular. The theoretical basis is that minimizing information entropy corresponds to the lowest uncertainty of the signal, such as... Figure 4 As shown by the convergence curve, the algorithm converges quickly and outputs the globally optimal parameter combination. .

[0077] S22: Decomposition and Complexity Measurement

[0078] Using the optimal parameters obtained in step S21, the original wind power sequence is decomposed using VMD to obtain K IMF components. To quantify its dynamic characteristics, the sample entropy of each IMFk is calculated. Sample entropy is a robust metric for measuring the complexity and irregularity of time series data; a higher entropy value indicates a more complex and unpredictable sequence. The embedding dimension m for sample entropy calculation is set to 2.

[0079] S23: Complexity-based clustering and recombination

[0080] Using sample entropy as a feature, the K-Means clustering algorithm is employed for unsupervised partitioning of all IMF components. The number of clusters (C) is dynamically determined using metrics such as silhouette coefficient; in this embodiment, the optimal number of clusters is 3. Figure 3 As shown, based on the clustering results, the IMF is reorganized into three feature subsets with clearly defined physical meanings:

[0081] S231: Low-complexity subset This corresponds to the minimum entropy class, representing the deterministic global trend of the sequence;

[0082] S232: Medium-complexity subset This corresponds to the class with a medium entropy value, representing a regular fluctuation component;

[0083] S233: Highly Complex Subset This corresponds to the class with the largest entropy value, representing the randomness and local transient nature of the sequence.

[0084] This process enables data-driven, downstream-oriented differentiation feature guidance for modular units.

[0085] S3: Dual-stream feature extraction and dynamic gating fusion Figure 5 (This demonstrates its internal mechanisms)

[0086] S31: Parallel Extraction of Heterogeneous Features

[0087] S311: Global feature extraction. For low-complexity subsets. A global modeling network based on a probabilistic sparse attention mechanism is adopted. This mechanism efficiently extracts global dependency features across long time series by significantly reducing the computational complexity of attention. :

[0088]

[0089] in, For global network parameters, hollow R ( ) is used to represent the set of real numbers. Representing a dimension as The space of real matrices; The time step (or sequence length) of the input time series. This represents the dimension of the feature channel.

[0090] S312: Local feature extraction. For high-complexity subsets. A bidirectional long short-term memory network is used. Processing is performed to capture its forward and backward local contextual information to obtain features. .

[0091] S313: Dimension Alignment. To fuse with global features in a unified space, a linear projection layer is introduced. Perform dimension alignment:

[0092]

[0093] in, This represents the original local feature matrix output by the local feature extraction network; Weights and biases for dimension alignment layers.

[0094] S32: Dynamic Gated Adaptive Fusion

[0095] S321: Feature splicing. (The sentence is incomplete and likely refers to a specific step in the process.) and The concatenation is performed on the last dimension to obtain the joint context representation. .

[0096] S322: Generate dynamic weights. Combine contextual features. The input is processed by a dynamic gating network consisting of a multilayer perceptron (MLP) and a sigmoid function, generating a dynamic weight matrix with the same dimension as the joint context features. :

[0097]

[0098] in, Use the Sigmoid activation function; , The weights and biases of the gated network; Represents the dynamic weight matrix The Middle The time step, the first The element values ​​of each feature dimension, due to the mapping of the Sigmoid function, are strictly within the range of (0, 1).

[0099] S323: Weighted fusion. Utilizing G pairs and By performing element-wise weighted summation, we obtain the fusion feature F:

[0100] .

[0101] This model achieves dual adaptation in both time step and feature dimension. When the sequence is stationary, G→1, and the model depends on the global trend. When the sequence undergoes a dramatic change, G→0, and the model focuses on local transients. .in, To be related to the dynamic weight matrix A matrix of all dimensions consisting entirely of 1s, meaning that all elements in the matrix have a value of 1.

[0102] S324: Preliminary prediction. The preliminary prediction result is obtained by linear projection of the fused feature F. .

[0103] S4: Frequency Domain Closed-Loop Error Correction and Joint Training

[0104] S41: Frequency Domain Transformation and Learnable Complex Filtering

[0105] S411: Frequency domain conversion. Preliminary predictions will be made. Mapping to the frequency domain using Real Fast Fourier Transform (RFFT): ,in, Represents the operator for the real number Fast Fourier Transform; The length is the frequency domain length. for Complex space; hollow It is specifically used to represent the set of complex numbers.

[0106] S412: Complex Filtering. To correct frequency domain distortion, this invention introduces a learnable complex filter. Each of its elements model and argument These are used to adjust the amplitude and phase of the corresponding frequency components, respectively. Dot-multiplication filtering is performed in the frequency domain:

[0107]

[0108] This operation is equivalent in the time domain to a non-causal complex convolution, which can simultaneously compensate for gain and time delay in the waveform. Represents the filtered and corrected spectrum (or frequency domain prediction result); symbol The dot product operator represents element-wise multiplication. This is the preliminary predicted frequency domain mapping result; It is a learnable complex filter matrix.

[0109] S42: Temporal Reconstruction and Closed-Loop Feedback

[0110] S421: Inverse Transform. Perform an inverse real fast Fourier transform (IRFFT) on the corrected spectrum and take the real part to obtain the time-domain correction term Δ:

[0111]

[0112] in, The time-domain correction term obtained from the transformation; operator Represents the Inverse Fast Fourier Transform (IRFFT) operation; operator This represents the mathematical operation of extracting the real part of a complex number.

[0113] S422: Residual connection. Final prediction result. By using a learnable scaling factor The residual connection yields:

[0114]

[0115] in, It is initialized to 0.1 and optimized during training to act as a damper for controlling the correction strength.

[0116] S43: End-to-end Joint Training

[0117] The parameters of the entire system are optimized using a joint loss function:

[0118]

[0119] Where Y is the true value, β is the weight coefficient, and the value of β ranges from 0 to 1. The reason for the value of β is as follows: β is used to adjust the contribution of the final prediction main loss, so it should be positive; β=0 will make the main loss ineffective, and β<0 does not conform to the loss minimization logic; too large a value of β can weaken the supervisory role of the auxiliary loss on the previous network and disrupt the joint training balance; limiting β to 0 to 1 is conducive to balancing the dominance of the main loss and the synergy of the auxiliary loss, and improving training stability. This is an initial estimate of the auxiliary loss; This represents the final predicted main loss.

[0120] In this embodiment, The value is set to 1.0. This loss forces the backbone network to generate error modes that are "easily corrected by frequency domain filters," while simultaneously driving the complex filter W. freq By studying the spectral evolution of system errors, a closed-loop learning system from feature extraction to error compensation is formed, realizing the mechanistic modeling of error dynamics.

[0121] S5: Comparative Experiments and Ablation Analysis

[0122] To fully verify the effectiveness of each module of the present invention, the following comparative experiments and ablation analyses were designed.

[0123] S51: Evaluation Indicators

[0124] To objectively quantify the predictive performance of the model, five standard evaluation metrics are used, including Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²).

[0125] S52: Baseline Model Comparison

[0126] Autoformer was selected as the baseline model for comparison. Autoformer is a variant of Transformer based on sequence decomposition and autocorrelation mechanisms, and is representative in the field of time series forecasting.

[0127] As shown in the comparison results, the CAG-DFC-Net proposed in this invention has achieved significant improvements in all core metrics, as shown in Table 1:

[0128] Table 1

[0129] Model MSE MAE R² Autoformer 0.1563 0.3209 -0.9408 CAG-DFC-Net (This Invention) 0.0036 0.0452 0.9558 Increase The decrease was 97.70%. The decrease was 85.91%. Increased to 0.9558

[0130] The baseline Autoformer model suffers from a high MSE of 0.1563 and a negative R² (-0.9408), indicating that the traditional Transformer architecture based on sequence decomposition exhibits severe feature extraction failure and error accumulation when dealing with time-series data such as wind power, which is characterized by high-frequency abrupt changes and strong non-stationarity. In contrast, CAG-DFC-Net significantly reduces the MSE to 0.0036 (a reduction of 97.70%), the MAE to 0.0452 (a reduction of 85.91%), and achieves an R² of 0.9558. This breakthrough performance improvement demonstrates that the Dynamic Gated Fusion (DFC) mechanism and the Frequency Domain Closed-Loop Residual Correction (CRC) module designed in this invention can effectively solve the coordination problem of heterogeneous features of multiple variables and successfully compensate for high-frequency residuals in the frequency domain, greatly improving the model's ability to fit non-stationary time series.

[0131] S53: Training Dynamics Analysis

[0132] The training dynamics of the baseline Autoformer model over 30 epochs. Training log data shows that the model exhibited a normal convergence trend in the first 15 epochs, with the training set MSE steadily decreasing from 1.05. However, around epoch 17 (after the test MSE reached a minimum of 0.0838), the model began to encounter a generalization bottleneck. Subsequently, although the training set error continued to decrease slowly (eventually converging to around 0.027), the test set error stagnated or even slightly rebounded (stabilizing at around 0.10).

[0133] This typical phenomenon of "underfitting and limited generalization ability" reveals the limitations of traditional Autoformer models: their internal autocorrelation mechanism relies excessively on the inherent long-range periodicity of historical sequences. When faced with wind power sequences, which are affected by extreme weather conditions and exhibit numerous random transients, relying solely on moving average sequence decomposition in the time domain cannot effectively extract physically meaningful trend terms, resulting in a significant reduction in the model's predictive performance on unknown test sets. This also provides strong counter-argument for introducing a Local Feature Extractor (BiLSTM) in this invention.

[0134] S54: Comparative Analysis of Model Convergence Performance

[0135] S541: Convergence Performance Analysis of the Model of this Invention

[0136] Please see Figure 6 , Figure 6This figure illustrates the training convergence process of the CAG-DFC-Net neural network model proposed in this invention on a single-step prediction task. The horizontal axis represents the number of iterations (Epochs), and the vertical axis represents the mean squared error loss (Loss). The blue dotted lines represent the training set loss (Train MSE loss), and the red asterisked lines represent the independent test set loss (Test MSE loss).

[0137] As shown in the figure, the training loss and test loss of the model in this invention decrease rapidly and synchronously from the initial stage, converging and stabilizing at an extremely low level close to 0.000 after approximately 10 training epochs. The two curves highly overlap throughout the process, ultimately reaching the same limit value. This result demonstrates that the model in this invention achieves fast, stable, and consistent convergence, exhibiting excellent generalization ability while achieving extremely high prediction accuracy, effectively overcoming the overfitting problem.

[0138] S542: Convergence Performance Analysis of Comparative Models

[0139] Please see Figure 7 , Figure 7 The training convergence process of the Autoformer model, which is the closest to existing technology, is demonstrated under the same task and dataset for comparison with this invention. Coordinate settings are the same. Figure 6 .

[0140] As shown in the figure, the training loss of this comparative model plateaus after rapidly decreasing to approximately 0.2, making further optimization impossible. While its test loss drops to 0.0, a significant "performance gap" emerges between it and the stagnant training loss. This result reveals the inherent flaws of existing models in such tasks: insufficient convergence and inconsistent training-test performance. Specifically, the models are prone to getting trapped in local optima, and their generalization performance lacks stable support from the training process.

[0141] S543: Summary of Convergence Performance Comparison

[0142] contrast Figure 6 and Figure 7 As can be seen, the CAG-DFC-Net model proposed in this invention significantly outperforms the existing Autoformer model in terms of convergence speed, convergence accuracy, and training-test consistency. These advantages are attributed to the synergistic effect of the three core modules designed in this invention: the CAG module provides high-quality input features, the DFC module implements adaptive feature fusion, and the CRC module performs closed-loop correction of errors in the frequency domain. These three modules are deeply coupled through end-to-end joint training, jointly ensuring the model's rapid convergence and excellent generalization ability.

[0143] S55: Comparison of Predictive Curves

[0144] To more intuitively demonstrate the predictive behavior characteristics of the model, prediction results from 150 consecutive time steps were extracted from the test set for visualization and comparison. The results are as follows: Figure 2 The dark solid line represents the actual normalized wind power fluctuation, which exhibits obvious multi-scale non-stationary characteristics.

[0145] Baseline model performance (dashed line): While Autoformer roughly captures the low-frequency periodic phase of wind power (i.e., the approximate location of peaks and troughs), it exhibits significant amplitude amplification (overshooting) distortion. This is because its multivariate decoder forcibly superimposes trend interference from non-target variables during the initialization phase and lacks the ability to constrain high-frequency residuals, causing the predicted curve to deviate from the true trajectory like a runaway horse.

[0146] The model's performance (light solid line): The CAG-DFC-Net's prediction curve closely wraps around and tracks the trajectory of actual wind power. Even in power abrupt changes with large gradient increases / decreases, such as in steps 40 and 85, the model can still follow smoothly and accurately. This is thanks to our designed frequency domain closed-loop error correction module (CRC). This module maps the initial open-loop prediction results of the backbone network to the frequency domain, accurately filters out phase drift and periodic deviations caused by heterogeneous feature fusion through a learnable complex filter, and finally superimposes the compensated residuals back into the time domain through inverse transformation, thus achieving high-fidelity disturbance-resistant prediction with the level of a "closed-loop control system".

[0147] S56: Ablation Test

[0148] To verify the effectiveness of each module of this invention, the following ablation model was designed:

[0149] Ours w / o CAG: Remove adaptive VMD recombination and replace it with fixed parameter (K=8, α=2000) VMD, and all IMF inputs are a single BiLSTM.

[0150] Ours w / o DFC: Remove dynamic gating, H g and H l Linear mapping after direct concatenation.

[0151] Ours w / o CRC: Removes frequency domain closed-loop correction, only outputs .

[0152] Ablation experiments show that each module makes a significant contribution to the final performance, with the CRC module showing the most significant improvement in the accuracy of medium- and long-term predictions, thus verifying the effectiveness of frequency domain closed-loop correction.

[0153] Example 2: Wind Power Prediction System

[0154] This embodiment provides a wind power prediction system based on time-frequency coordination and adaptive decoupling, including:

[0155] The data preprocessing module is used to acquire and standardize historical wind power series and related meteorological covariate data;

[0156] The Complexity Aware Decoupling Module (CAG Module) is used to perform adaptive mode decomposition and complexity-based recombination on the historical wind power data. This module includes: a parameter optimization unit, used to optimize VMD parameters with the goal of minimizing the average envelope entropy of the decomposed components; a decomposition and quantization unit, used to perform VMD decomposition and calculate the sample entropy of each component; and a clustering and recombination unit, used to cluster the components based on the sample entropy and recombine them into low-complexity, medium-complexity, and high-complexity subsets.

[0157] The Dynamic Feature Fusion Prediction Module (DFC Module) includes: a heterogeneous feature extraction submodule, used to extract features from the recombined sub-signal set, including a global feature extractor based on an attention mechanism and a local feature extractor based on a recurrent neural network; and a dynamic gated fusion submodule, used to adaptively fuse the extracted heterogeneous features and output preliminary power prediction results.

[0158] The frequency domain closed-loop correction module (CRC module) is used to filter and correct the error of the preliminary power prediction result in the frequency domain, and reconstruct the final power prediction result.

[0159] The model training module is used to collaboratively optimize the parameters of each module through an end-to-end joint loss function.

[0160] For the specific implementation methods of the above modules, please refer to the corresponding steps in Example 1, which will not be repeated here.

[0161] This wind power prediction method and system, based on time-frequency collaboration and adaptive decoupling, optimizes VMD parameters through "minimum average envelope entropy" and combines sample entropy clustering and recombination to achieve accurate decoupling of non-stationary sequences for prediction tasks, overcoming the shortcomings of parameter empiricalization and homogenized processing. The dynamic gating mechanism enables context-aware and adaptive modulation of global and local feature contributions, improving the model's adaptability to different operating conditions and feature utilization efficiency. Frequency-domain complex filtering correction can simultaneously correct the amplitude and phase errors of the prediction results, and through closed-loop joint training, it suppresses the accumulation of errors in multi-step predictions from a mechanistic perspective, significantly improving medium- and long-term prediction accuracy. The three modules work in deep collaboration under end-to-end training, breaking the defective propagation chain of the traditional linear paradigm and achieving optimal global performance.

[0162] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These will not affect the effectiveness and practicality of the implementation of the present invention.

Claims

1. A wind power prediction method based on time-frequency coordination and adaptive decoupling, characterized in that, Includes the following steps: S1. Obtain historical wind power data; S2. Perform complexity-aware adaptive mode decomposition and recombination on the historical wind power data to obtain multiple sets of reconstructed sub-signals with different dynamic complexity characteristics; S3. Input the multiple reconstructed sub-signal sets into a heterogeneous feature extraction network for parallel feature extraction to obtain heterogeneous features; S4. Adaptively weighted fusion of the heterogeneous features is performed through a dynamic gated fusion network to generate fused features, and preliminary power prediction results are obtained based on the fused features. S5. Input the preliminary power prediction result into the frequency domain closed-loop correction network, correct the spectrum of the preliminary power prediction result through a learnable frequency domain filter, convert the corrected spectrum back to the time domain, and compensate the preliminary power prediction result in the form of residuals to obtain the final power prediction result. In particular, steps S2, S3, S4 and S5 are performed through collaborative training using an end-to-end joint loss function.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: S21. With minimizing the average envelope entropy of each component after decomposition as the optimization objective, adaptively determine the number of modes K and the penalty factor α of variational mode decomposition; S22. Using the determined number of modes K and the penalty factor α, variational mode decomposition is performed on the historical wind power data to obtain multiple intrinsic mode function components; S23. Calculate the sample entropy of each intrinsic mode function component, and perform unsupervised clustering and recombination on all components based on the sample entropy to form the reconstructed sub-signal set, which includes at least a low-complexity subset and a high-complexity subset.

3. The method according to claim 2, characterized in that, Step S3 specifically includes: S31. Input the low-complexity subset into the global feature extractor to obtain long-range dependency features; S32. Input the high-complexity subset into the local feature extractor to obtain local transient features; S33. Align the dimensions of the local transient features with those of the long-range dependent features.

4. The method according to claim 3, characterized in that, Step S4 specifically includes: S41. The long-range dependency features are concatenated with the dimension-aligned local transient features to obtain joint context features; S42. Input the joint context features into a gating network to generate a dynamic weight matrix G with the same dimension as the joint context features. The output range of the gating network is (0, 1). S43. Using the dynamic weight matrix, perform element-wise weighted summation on the long-range dependency feature and the local transient feature to obtain the fused feature F: , where H g H is a long-range dependency feature. l These are local transient features after dimension alignment. This represents element-wise multiplication. To be related to the dynamic weight matrix A matrix of all dimensions consisting entirely of 1s, meaning that all elements in the matrix have a value of 1.

5. The method according to claim 3, characterized in that, Step S5 specifically includes: S51. Perform a real-number fast Fourier transform on the preliminary power prediction result and map it to the frequency domain to obtain the frequency domain prediction result. S52. Multiply the frequency domain prediction result element-wise with the learnable complex filter to obtain the corrected spectrum. Each element of the complex filter is used to adjust the amplitude and phase of the corresponding frequency component. S53. Perform a real-number fast Fourier inverse transform on the corrected spectrum and take the real part to obtain the time-domain correction term; S54. Add the time-domain correction term to the preliminary power prediction result to obtain the final power prediction result.

6. The method according to claim 5, characterized in that, Step S54 specifically involves: ； in, For the final power prediction results, This is the preliminary power prediction result, where Δ is the time-domain correction term. This is a learnable scaling factor.

7. The method according to claim 1, characterized in that, The end-to-end joint loss function The final power prediction result is weighted and consists of a first loss between the initial power prediction result and the actual value, and a second loss between the final power prediction result and the actual value. ； Where Y is the true value, β is the weighting coefficient, and the value of β ranges from 0 to 1. For preliminary prediction of auxiliary losses, This represents the final predicted main loss.

8. A wind power prediction system based on time-frequency coordination and adaptive decoupling, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data preprocessing module is used to acquire and standardize the historical wind power sequence; A complexity-aware decoupling module is used to perform adaptive mode decomposition and complexity-based recombination on the standardized wind power sequence; The dynamic feature fusion prediction module includes: a heterogeneous feature extraction submodule, used to extract features from the recombined sub-signal set; and a dynamic gated fusion submodule, used to adaptively fuse the extracted heterogeneous features and output preliminary power prediction results. The frequency domain closed-loop correction module is used to filter and correct the error of the preliminary power prediction result in the frequency domain, and reconstruct the final power prediction result. The collaborative training module is used to collaboratively optimize the parameters of the complexity-aware decoupling module, the dynamic feature fusion prediction module, and the frequency domain closed-loop correction module through an end-to-end joint loss function.

9. The wind power prediction system based on time-frequency coordination and adaptive decoupling according to claim 8, characterized in that, The complexity-aware decoupling module includes: The parameter optimization unit is used to optimize VMD parameters with the goal of minimizing the average envelope entropy of the decomposed components. The decomposition and quantization unit is used to perform VMD decomposition and calculate the sample entropy of each component; Clustering and recombination units are used to cluster components based on sample entropy and recombine them into multiple reconstructed subsets, including at least low-complexity subsets and high-complexity subsets.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

Images