A wind power prediction method based on residual correction and related equipment

Abstract

The present application belongs to the technical field of wind power generation, and particularly relates to a wind power prediction method based on residual correction and related equipment. A historical wind power sequence is obtained, and a plurality of intrinsic mode components are obtained through complete set total experience mode decomposition primary decomposition. The sample entropy of each component is calculated and classified into noise, volatility and trend items. After removing the noise items, the volatility items are twice decomposed into quasi-stationary sub-sequences through variational mode decomposition. The quasi-stationary sub-sequences and the trend items are extracted for feature construction to form an enhanced feature matrix through fuzzy information granulation. The enhanced feature matrix is input into a dynamic weighted multi-kernel support vector regression model, and the parameters and kernel weights are optimized by an improved nuthatch algorithm. The preliminary prediction residual is calculated, the correction increment is generated by using a time convolution network, and the final prediction value is reconstructed by combining the preliminary prediction value. The present application suppresses mode aliasing through cascading decomposition, multi-kernel regression and residual closed-loop correction, and improves the prediction accuracy of non-stationary sequences.

Description

Technical Field

[0001] This invention belongs to the field of wind power generation technology, specifically relating to a wind power prediction method and related equipment based on residual correction. Background Technology

[0002] Wind power output exhibits significant randomness and non-stationarity, with an approximately cubic relationship between power and wind speed. Small deviations in wind speed forecasting can be drastically amplified at the power level, leading to inaccurate dispatching plans and threatening the safe and stable operation of the power grid. As the installed capacity of wind power continues to expand and the proportion of wind power connected to the grid continues to increase, the requirements for the accuracy and reliability of wind power forecasting are becoming increasingly stringent.

[0003] In wind power forecasting, Support Vector Machines (SVMs) are widely used due to their advantages in handling nonlinear regression problems. However, the prediction accuracy of SVMs is highly dependent on the selection of hyperparameters such as kernel function parameters and penalty factors; inappropriate parameter selection will significantly reduce model performance. To optimize SVM parameters, traditional Particle Swarm Optimization (PSO) and Grey Wolf Optimization (GWO) algorithms are often used for parameter optimization, but these algorithms are prone to getting trapped in local optima in the later stages of the search, making it difficult to obtain the globally optimal parameter combination. Furthermore, single-core SVMs struggle to simultaneously adapt to the coexisting low-frequency trend components and high-frequency fluctuation components in wind power data, resulting in insufficient fitting ability for complex wind power sequences.

[0004] To improve prediction accuracy, signal decomposition methods have been introduced into the field of wind power forecasting. Methods such as Empirical Mode Decomposition (EMD) can decompose the original wind power sequence into multiple intrinsic mode functions (EMFs). However, EMD suffers from severe mode aliasing, the physical meaning of the decomposed components is ambiguous, and boundary effects can affect prediction stability. In recent years, deep learning models such as Long Short-Term Memory (LSTM) networks have also been applied to wind power forecasting. However, when faced with non-stationary residuals caused by meteorological abrupt changes, LSTMs often exhibit prediction response lag, making it difficult to quickly track the abrupt changes in wind power.

[0005] In summary, existing wind power prediction technologies suffer from the following shortcomings: SVM parameter optimization is prone to getting trapped in local optima; single-core SVM struggles to accommodate multi-scale features; signal decomposition methods suffer from mode aliasing; and deep learning models exhibit lag in response to non-stationary residuals. Therefore, there is an urgent need for a wind power prediction scheme capable of deeply deconstructing complex signals, possessing global parameter optimization capabilities, and implementing closed-loop error correction to overcome the aforementioned deficiencies of existing technologies. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a wind power prediction method and related equipment based on residual correction, which addresses the shortcomings of the prior art. This method solves the technical problem that traditional particle swarm optimization and gray wolf optimization are prone to getting trapped in local optima in the later stages of the search and are difficult to obtain the global optimal parameter combination.

[0007] The objective of this invention is achieved through the following technical solutions: In a first aspect, the present invention provides a wind power prediction method based on residual correction, comprising: Historical wind power sequences are obtained, and the historical wind power sequences are decomposed into several intrinsic mode function components using the fully lumped empirical mode decomposition technique. Calculate the sample entropy value of each intrinsic mode function component, and classify each intrinsic mode function component into noise term, volatile term and trend term based on the sample entropy value; The noise term is removed, and the volatile term is decomposed a second time using the variational mode decomposition algorithm to obtain a quasi-stationary subsequence; The quasi-stationary subsequence and the trend term are feature extracted using fuzzy information granulation technology to construct an enhanced feature matrix; The enhanced feature matrix is ​​input into the dynamic weighted multi-kernel support vector regression model for prediction, and the preliminary prediction value is output. The model parameters and kernel weights of the dynamic weighted multi-kernel support vector regression model are obtained by optimizing the improved Super Flycatcher algorithm. Calculate the residual sequence of the preliminary prediction value, use a temporal convolutional network to generate a correction increment based on the residual sequence, reconstruct the correction increment with the preliminary prediction value, and output the final wind power prediction value.

[0008] As a further improvement of the present invention, the primary decomposition of the historical wind power sequence using the fully lumped empirical mode decomposition technique yields several intrinsic mode function components, including: For multiple decomposition experiments, a standard normally distributed white noise sequence is added to the historical wind power sequence each time, and the sequence is decomposed to extract the corresponding temporary mode components. Calculate the ensemble average of the temporary modal components obtained from multiple decomposition experiments, and use the ensemble average as the eigenmode function component of the corresponding order.

[0009] As a further improvement of the present invention, the step of classifying each intrinsic mode function component into noise term, volatile term, and trend term according to the sample entropy value includes: A first entropy threshold and a second entropy threshold are preset, wherein the first entropy threshold is less than the second entropy threshold; Compare the sample entropy value with the first entropy threshold and the second entropy threshold: If the sample entropy value is less than the first entropy threshold, then the corresponding intrinsic mode function component is classified as the trend term; If the sample entropy value is greater than or equal to the first entropy threshold and less than the second entropy threshold, then the corresponding intrinsic mode function component is classified as the volatile term; If the sample entropy value is greater than or equal to the second entropy threshold, then the corresponding intrinsic mode function component is classified as the noise term.

[0010] As a further improvement of the present invention, the step of performing a secondary decomposition of the volatile term using a variational mode decomposition algorithm to obtain a quasi-stationary subsequence includes: A variational optimization model for the volatile term is constructed with the goal of minimizing the sum of the estimated bandwidths of each mode. The volatile term is solved iteratively using the alternating direction multiplier method with the variational optimization model to separate several frequency band-limited modal components, which are then used as the quasi-stationary subsequence.

[0011] As a further improvement of the present invention, the step of extracting features from the quasi-stationary subsequence and the trend term using fuzzy information granulation technology to construct an enhanced feature matrix includes: The quasi-stationary subsequence and the trend term are divided into sliding blocks according to a preset time window; Fuzzy membership functions are used to map data points within each time window, and the low-order extreme value, mean value, and high-order extreme value of the data within each time window are calculated and extracted. The low-order extreme values, mean values, and high-order extreme values ​​corresponding to each time window are combined and concatenated to construct the enhanced feature matrix.

[0012] As a further improvement of the present invention, the combined kernel function in the dynamically weighted multi-kernel support vector regression model is a linear weighted sum of at least two basis kernel functions; The weight coefficients of each of the basis kernel functions are dynamic weighting coefficients, and the sum of the dynamic weighting coefficients of all basis kernel functions is equal to 1.

[0013] As a further improvement of the present invention, the model parameters and kernel weights of the dynamically weighted multi-kernel support vector regression model are obtained by optimizing the improved Super Flycatcher algorithm, including: The improved super flycatcher algorithm updates the individual positions of the population using an individual position update equation that includes a Levy flight perturbation term during the iterative optimization process. The optimal individual position obtained after optimization convergence is analyzed and mapped to the penalty factor of the dynamically weighted multi-kernel support vector regression model and the dynamic weighting coefficient of each basis kernel function.

[0014] As a further improvement of the present invention, the step of generating a correction increment based on the residual sequence using a temporal convolutional network includes: The residual sequence is input into the dilated convolutional layer of the temporal convolutional network to extract nonlinear error features from the time series; The nonlinear error features are mapped and output through the fully connected layer of the temporal convolutional network to obtain the correction increment for the corresponding time step.

[0015] Secondly, the present invention provides a wind power prediction system based on residual correction, comprising: The primary decomposition module is used to obtain the historical wind power sequence and perform primary decomposition on the historical wind power sequence using the fully lumped empirical mode decomposition technique to obtain several intrinsic mode function components. The complexity classification module is used to calculate the sample entropy value of each intrinsic mode function component and classify each intrinsic mode function component into noise, volatility and trend terms based on the sample entropy value. The secondary decomposition module is used to remove the noise term and perform secondary decomposition on the volatile term using the variational mode decomposition algorithm to obtain a quasi-stationary subsequence. The feature enhancement module is used to extract features from the quasi-stationary subsequence and the trend term using fuzzy information granulation technology, and construct an enhanced feature matrix; The multi-core prediction module is used to input the enhanced feature matrix into the dynamic weighted multi-core support vector regression model for prediction and output preliminary prediction values. The model parameters and kernel weights of the dynamic weighted multi-core support vector regression model are obtained by optimizing the improved Super Flycatcher algorithm. The residual correction module is used to calculate the residual sequence of the preliminary prediction value, generate a correction increment based on the residual sequence using a temporal convolutional network, reconstruct the correction increment with the preliminary prediction value, and output the final wind power prediction value.

[0016] Thirdly, the present invention provides a computer device, comprising: a processor and a computer-readable storage medium; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program that, when executed by the processor, implements the wind power prediction method based on residual correction as described above.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed as described above for wind power prediction based on residual correction.

[0018] The beneficial effects of this invention are as follows: This invention provides a wind power prediction method based on residual correction. It utilizes fully lumped empirical mode decomposition (EMD) technology to perform a primary decomposition of historical wind power sequences to obtain several intrinsic mode function (EMF) components. By adding paired positive and negative auxiliary noise to the original signal and canceling residual noise in the ensemble averaging, the inherent mode aliasing phenomenon of EMD is suppressed, making the spectral boundaries of each component clearer. The sample entropy value of each EMF component is calculated, and the components are classified into noise, volatile, and trend terms accordingly. The sample entropy quantifies the complexity and irregularity of each component, achieving objective classification based on information entropy and avoiding classification bias caused by subjective threshold settings. The model is designed to improve performance. Noise terms are removed, and the volatile term is further decomposed using variational mode decomposition to obtain a quasi-stationary subsequence. This process filters out high-frequency noise interference and further decomposes non-stationary fluctuation components into quasi-stationary subsequences with compact spectra and well-defined center frequencies, reducing the fitting difficulty of subsequent prediction models. Fuzzy information granulation technology is used to extract features from the quasi-stationary subsequences and trend terms, constructing an enhanced feature matrix. This transforms the original numerical sequence into fuzzy information granules containing lower bounds, kernels, and upper bounds, extracting interval and trend features of the sequence and enhancing the ability of the input features to represent the boundaries of wind power fluctuations. The enhanced feature matrix is ​​then input into a dynamically weighted multi-kernel support vector regression model for prediction and outputs the initial prediction result. The model uses different kernel functions to map different frequency domain features, and dynamic weighting allows low-frequency trends and high-frequency fluctuations to obtain suitable mapping channels, avoiding the limitation of single-kernel support vector machines in taking into account multi-scale features. Furthermore, the model parameters and kernel weights are optimized using an improved Super Flying Maneuver algorithm. This algorithm introduces chaotic mapping to initialize the population and incorporates the Lévy flight mechanism, expanding the population distribution range in the early search phase and enhancing the ability to escape local optima in the later search phase. It overcomes the premature convergence defects of traditional particle swarm optimization and gray wolf optimization, obtaining a globally better combination of model parameters and kernel weights. The residual sequence of the initial predicted values ​​is calculated, and a temporal convolutional network is used to analyze the residual values. The difference sequence generates a correction increment and reconstructs the final wind power prediction value with the preliminary prediction value. The temporal convolutional network captures the long-term temporal dependence in the residual sequence through causal convolution and dilatational convolution, and quickly responds to and compensates for the non-stationary residuals caused by meteorological changes in the preliminary prediction, overcoming the problem of the long short-term memory network's lagging response to abrupt residuals. The cascaded decomposition of this invention removes noise and reduces non-stationarity at the signal level, fuzzy information granulation enhances boundary representation at the feature level, dynamic weighted multi-kernel support vector regression realizes multi-scale mapping at the model level, and the temporal convolutional network closes the loop to correct abrupt deviations at the error level. The collaboration of each level ensures that the prediction results maintain high accuracy in both stationary and abrupt periods. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the wind power prediction method based on residual correction in an embodiment of the present invention. Figure 2 This is an internal structural diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives and technical solutions of this invention clearer and easier to understand, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0023] The dispatchability of wind power output is highly dependent on the accuracy of power prediction. However, wind power sequences are characterized by nonlinearity, non-stationarity, and multi-scale coupling, which poses a continuous technical challenge to accurate prediction.

[0024] First, at the level of original signal decomposition, traditional empirical mode decomposition methods generally suffer from mode aliasing during component extraction. That is, oscillating components at different time scales are incorrectly assigned to the same intrinsic mode function, making it difficult to separate trend changes from high-frequency random fluctuations in the subsequent modeling stage. Even with the improved form of lumped empirical mode decomposition, there are still shortcomings such as residual noise not being completely eliminated and overlapping frequency bands of adjacent components, which limits the physical interpretability of the decomposition results.

[0025] Secondly, at the modeling algorithm level, support vector regression has theoretical advantages in small-sample nonlinear regression tasks, but its performance is highly dependent on the choice of penalty factor and kernel function parameters. Single-kernel support vector regression struggles to balance global fitting capability and local sensitivity when faced with composite wind power signals exhibiting both low-frequency, gradual trends and high-frequency, abrupt fluctuations. Traditional heuristic optimization methods such as particle swarm optimization and gray wolf algorithms are prone to getting trapped in local optima in the later stages of iteration, making it difficult to provide stable and reliable global parameters for multi-kernel models.

[0026] Third, in terms of residual processing, when relying solely on the single-stage regression model to output prediction results, the model's response to inflection point conditions such as sudden weather changes and rapid increases in wind speed often lags behind the measured values. The nonlinear error evolution law contained in the residuals is not effectively utilized, and the prediction accuracy decreases significantly under extreme conditions.

[0027] Fourth, at the feature representation level, the point-by-point features obtained by directly sampling the power sequence are easily contaminated by microscopic noise caused by wind field turbulence. The model is not robust enough to short-term extreme fluctuations and it is difficult to stably output high-quality predicted values ​​that can be used by the scheduling system.

[0028] To address the aforementioned shortcomings, this application provides a wind power prediction method based on residual correction. By employing complexity-aware multi-level signal deconstruction, feature enhancement through fuzzy information granulation, global optimization of the dynamically weighted multi-kernel support vector regression model using an improved Super Flying Man algorithm, and closed-loop residual compensation using a temporal convolutional network, a high-precision prediction link that can be deployed in an engineering manner is formed.

[0029] Example 1 This embodiment provides a wind power prediction method based on residual correction, such as... Figure 1 As shown. The following is a detailed implementation method.

[0030] S1. Obtain historical wind power sequences and perform primary decomposition of the historical wind power sequences using the fully lumped empirical mode decomposition technique to obtain several intrinsic mode function components.

[0031] Specifically, the historical wind power sequence refers to the active power time series continuously recorded by the data acquisition and SCADA monitoring and control system of a wind farm or wind farm group over a continuous period of time in the past. The time series variable is labeled as P(t), where t represents the time series sampling time step, and the sampling interval can be set to 15 minutes according to the ultra-short-term dispatch requirements. In this embodiment, the fully lumped empirical mode decomposition (EMD) technique is a multi-scale decomposition method for non-stationary nonlinear signals. Its working mechanism involves repeatedly adding controlled-amplitude standard normally distributed white noise to the original signal, performing EMD on the composite signal after each addition, and taking a ensemble average of the same-order mode components obtained from multiple experiments to offset the influence of the added white noise, thereby suppressing the mode aliasing phenomenon present in traditional EMD. This technique preserves the energy distribution of the original signal while making the decomposition results more regular in frequency band division.

[0032] The intrinsic mode function components are several orders of subsequences output by the fully lumped empirical mode decomposition. Each order is distinct in frequency band and has a local narrow band in time domain, possessing independent physical meaning. After the power sequence is decomposed by the fully lumped empirical mode decomposition, it will be separated into multiple intrinsic mode function components corresponding to macroscopic trend changes, mid-frequency volatility, and high-frequency random disturbances.

[0033] Adaptive white noise is added during the decomposition process, and the ensemble average of each component is calculated to obtain the first-order eigenmode function. :

[0034]

[0035] Where N is the number of integrations. Let A be the modal component extracted in the i-th experiment. The modal aliasing problem is effectively solved by the above method, ensuring the integrity and orthogonality of the decomposed components. Here, A and B represent the template matching pairs of lengths m+1 and m under the tolerance r, respectively.

[0036] S2. Calculate the sample entropy value of each intrinsic mode function component, and classify each intrinsic mode function component into noise term, volatile term and trend term according to the sample entropy value.

[0037] Specifically, a first entropy threshold and a second entropy threshold are preset, and the first entropy threshold is less than the second entropy threshold; when the sample entropy value is less than the first entropy threshold, the component is identified as a trend term; when the sample entropy value is greater than or equal to the first entropy threshold and less than the second entropy threshold, the component is identified as a volatile term; when the sample entropy value is greater than or equal to the second entropy threshold, the component is identified as a noise term.

[0038] It should be noted that sample entropy is a statistical indicator that measures the complexity and irregularity of a time series. A higher value indicates that the sequence contains more random components and has lower predictability, while a lower value indicates that the sequence has a stronger trend regularity. This step calculates the sample entropy value for each eigenmode function component output by the complete set empirical mode decomposition, and quantifies and sorts each component in terms of complexity.

[0039] Noise components are high-frequency components with high sample entropy, strong dynamic randomness, and no significant gain on the final prediction accuracy; volatile components are mid-frequency components with moderate sample entropy, exhibiting both periodic fluctuations and random disturbances; trend components are low-frequency components with low sample entropy, evolving slowly over time and carrying the macroscopic evolution law of the power sequence. By classifying the intrinsic mode function components into these three categories, differentiated processing can be achieved, including noise removal, deep stabilization of volatile components, and direct retention of trend components.

[0040] S3. Remove noise terms and use variational mode decomposition algorithm to perform secondary decomposition on volatile terms to obtain quasi-stationary subsequences.

[0041] Specifically, with the goal of minimizing the sum of estimated bandwidths for each mode, a variational optimization model for the volatile term is constructed; the variational optimization model is used to iteratively solve the volatile term using the alternating direction multiplier method to separate several frequency band-limited modal components, which are then used as the quasi-stationary subsequence.

[0042] In this embodiment, noise removal refers to removing high-frequency components from the component set whose sample entropy value is higher than the second entropy threshold, whose contribution to the prediction target is low, and which will introduce additional modeling burden.

[0043] Variational mode decomposition (VMD) algorithms aim to minimize the sum of estimated bandwidths for each mode. They construct a constrained variational optimization model in the frequency domain and iteratively solve the model using the alternating direction multiplier method. This adaptively decomposes a composite signal into several bandwidth-limited modal components with clearly defined center frequencies. Compared to empirical mode decomposition methods, this algorithm possesses stronger mathematical rigor and avoids endpoint effects and modal overlap.

[0044] Quasi-stationary subsequences refer to several band-limited mode components output by the variational mode decomposition algorithm after applying it to the volatile term. These components approximate a stationary sequence, facilitating accurate fitting of its patterns by subsequent supervised learning models. The formula for calculating the quadratic decomposition is: This formula characterizes a constrained variational optimization model constructed for the volatile term. The objective function is to minimize the sum of the estimated bandwidths of each mode, and the constraint is that the sum of the modes is equal to the volatile term signal to be decomposed. This formula achieves deep stabilization of the mid-frequency volatile component, stripping the complex power volatile characteristics into a quasi-stationary signal.

[0045] S4. Use fuzzy information granulation technology to extract features from quasi-stationary subsequences and trend terms, and construct an enhanced feature matrix.

[0046] Specifically, fuzzy information granulation involves sliding blocks of the aligned stationary subsequence and trend term according to a preset time window, mapping the data points within each time window using fuzzy membership functions, calculating and extracting the low-order extreme values, mean values, and high-order extreme values ​​of the data within each time window, and combining and splicing the low-order extreme values, mean values, and high-order extreme values ​​corresponding to each time window to construct an enhanced feature matrix.

[0047] In this embodiment, a triangle or Gaussian membership function is used. The sub-components are granulated. Low, Avg, and High 3D features are extracted from each window to construct an enhanced feature matrix that can characterize the drastic changes in wind speed.

[0048] It should be noted that fuzzy information granulation technology refers to dividing a time series into sliding blocks using a preset time window, and within each time window, using a fuzzy membership function to map point-by-point data into information particles with flexible boundaries, thereby replacing point-by-point features with granular features in subsequent modeling. This technology can preserve the macroscopic fluctuation patterns within the window while filtering out microscopic noise disturbances caused by factors such as wind field turbulence. The enhanced feature matrix refers to a multidimensional feature matrix obtained by combining and concatenating the low-order extreme values, mean, and high-order extreme values ​​within each time window. This matrix uses the time window as the row dimension and the combination of low-order extreme values, mean, and high-order extreme values ​​as the column dimension, and can characterize the fluctuation amplitude and central trend of the wind power series at different scales, providing high-information-density input features for the support vector regression model.

[0049] S5. Input the enhanced feature matrix into the dynamic weighted multi-kernel support vector regression model for prediction, and output the preliminary predicted value. The model parameters and kernel weights of the dynamic weighted multi-kernel support vector regression model are obtained by optimizing the improved Super Fairy Flying Man algorithm.

[0050] In this embodiment, the dynamically weighted multi-kernel support vector regression model is a regression modeling framework based on the idea of ​​multi-kernel learning. Its core lies in constructing a combined kernel function by linearly weighting the sum of at least two basic kernel functions with different expressive capabilities. The weight coefficients of each basic kernel function are dynamically weighted coefficients, and the sum of the dynamic weighted coefficients of all basic kernel functions equals 1. Through the weighted combination of heterogeneous basic kernel functions, the model can simultaneously characterize the low-frequency smooth trend and high-frequency abrupt fluctuations in the wind power sequence, exhibiting stronger multi-scale modeling capabilities compared to single-kernel support vector regression.

[0051] Model parameters and kernel weights refer to the penalty factor and dynamic weighting coefficients of each basis kernel function in support vector regression. The penalty factor is used to balance training error and model complexity, while the dynamic weighting coefficients are used to distribute capabilities among multiple basis kernel functions. The values ​​of these two types of parameters significantly affect the model's fitting accuracy and generalization ability.

[0052] The improved Super Flycatcher algorithm is a swarm intelligence optimization algorithm that introduces chaotic mapping initialization and a Lévy flight perturbation term on the basis of the standard Super Flycatcher algorithm. The Lévy flight perturbation term, with its heavy-tailed step size mechanism, enables individuals in the population to escape local optima, allowing the iterative optimization process to cover a wider solution space. Finally, the optimal individual position obtained after optimization convergence is analyzed and mapped to the penalty factor of the dynamically weighted multi-kernel support vector regression model and the dynamic weighting coefficients of each basis kernel function.

[0053] Specifically, the DW-MKSVR model is constructed, and its combined kernel function is: The ISFOA algorithm is used to iteratively optimize the weights and SVR parameters. ISFOA introduces the Lévy flight update equation:

[0054] Ensure that the algorithm can escape local optima and obtain the globally optimal kernel parameter configuration.

[0055] Calculate the preliminary forecast value Extracting residual sequences The dilated convolution of TCN is used to capture the short-range nonlinearity in E(t), and the correction increment is calculated. Final output: .

[0056] S6. Calculate the residual sequence of the preliminary prediction values, use a time convolutional network to generate correction increments based on the residual sequence, reconstruct the correction increments with the preliminary prediction values, and output the final wind power prediction values.

[0057] Specifically, the residual sequence is input into the dilated convolutional layer of the temporal convolutional network to extract the nonlinear error features in the time series; the nonlinear error features are mapped and output through the fully connected layer of the temporal convolutional network to obtain the correction increment for the corresponding time step.

[0058] The following two examples illustrate the wind power prediction method based on residual correction of the present invention.

[0059] As an example, we use 15-minute ultra-short-term wind power prediction under normal and stable operating conditions in a plain wind farm. Specifically, we select historical SCADA data of normal and stable power generation throughout the year from wind farms in the central plains of China for prediction verification: SCADA data of a plain wind farm were collected, and the raw power P(t) was decomposed by CEEMDAN to produce 8 IMF components.

[0060] Calculate the sample entropy of each component. Among them, IMF1 =0.55, which was classified as a volatile term. Subsequently, the VMD formula was applied for secondary decomposition, which separated the high-frequency random fluctuations into 5 stationary subsequences.

[0061] Optimize DW-MKSVR parameters using ISFOA, where multi-core weights During the iteration process, the Lévy flight operator is used. It continuously escapes local optima. The final optimization result gives a Gaussian kernel weight of 0.72 and a linear kernel weight of 0.28.

[0062] At a 15-minute step size, using The calculated coefficient of determination is 0.985, and the MAPE error is as low as 3.8%. Plain wind farms experience no severe meteorological disturbances, and the time series exhibits a predominantly stable trend. This scheme relies on sample entropy grading to perform secondary decomposition only on a single volatile term, significantly reducing the computational cost of secondary decomposition of all components. Fuzzy information granular filtering filters out minor daily turbulent noise from the units. The multi-core model matches the stable trend with small short-term fluctuations, and TCN slightly corrects minor prediction biases under normal operating conditions. While ensuring ultra-high prediction accuracy, this approach shortens model training and inference time, adapting to the grid's normalized ultra-short-term dispatch requirements.

[0063] As an example, further, a dataset of sudden wind speed changes during cold waves at coastal tidal flat wind farms in eastern China was selected to conduct extreme condition verification: Power time series data were collected at 15-minute intervals over a 72-hour cold wave period. CEEMDAN primary decomposition was performed to obtain IMF components with multiple levels of mixed and violent fluctuations.

[0064] Calculate the entropy of each component sample. Components with multiple medium entropy values ​​are identified as volatile terms, high entropy value spikes are directly removed, and low entropy value long-term output trend terms are fully retained.

[0065] All volatile terms are uniformly subjected to VMD second variational decomposition, which splits them into multiple quasi-stationary subsequences and removes the violent short-term power oscillations caused by the cold wave.

[0066] During cold wave periods, the original time series contains a large number of numerical spikes caused by sudden changes in wind speed. Gaussian fuzzy membership functions are used for window granularization, and the three-dimensional features of the window are extracted to construct an enhanced feature matrix. This smooths out microscopic spikes while fully preserving the macroscopic fluctuation boundaries of sudden increases and decreases in wind speed.

[0067] The ISFOA algorithm is used to optimize the multi-core weights and hyperparameters of DW-MKSVR, and the Levy flight mechanism is used to avoid the optimization algorithm from getting stuck in local optima under drastic fluctuations, thus completing the initial prediction output. At the moment of abrupt change, the output of the DW-MKSVR model has a significant lag bias.

[0068] The residual sequence during the mutation period is extracted and fed into the TCN network. The expanded convolution accurately captures the mutation pattern of the residual and generates compensation increments, which are then superimposed on the preliminary prediction value to complete the closed-loop correction.

[0069] Test data from this embodiment show that the root mean square error (RMSE) of this solution is reduced by 67.08% compared to the traditional PSO-SVM prediction model.

[0070] The beneficial effects of this embodiment are as follows: For extreme weather conditions such as cold waves and gusts, CEEMD combined with VMD double-layer decomposition fully removes severe random fluctuations, and granularization of fuzzy information improves the model's robustness in identifying the characteristics of the mutation boundary; ISFOA global optimization ensures that the multi-core model adapts to the characteristics of the mutation scale; TCN network establishes a nonlinear compensation mechanism specifically for the lag residuals caused by mutations, which greatly reduces the prediction error of extreme conditions and improves the grid's dispatch safety margin in response to severe fluctuations in wind power output.

[0071] Example 2 This embodiment provides a wind power prediction system based on residual correction, specifically including: The primary decomposition module is used to obtain the historical wind power sequence and perform primary decomposition on the historical wind power sequence using the fully lumped empirical mode decomposition technique to obtain several intrinsic mode function components. The complexity classification module is used to calculate the sample entropy value of each intrinsic mode function component and classify each intrinsic mode function component into noise, volatility and trend terms based on the sample entropy value. The secondary decomposition module is used to remove the noise term and perform secondary decomposition on the volatile term using the variational mode decomposition algorithm to obtain a quasi-stationary subsequence. The feature enhancement module is used to extract features from the quasi-stationary subsequence and the trend term using fuzzy information granulation technology, and construct an enhanced feature matrix; The multi-core prediction module is used to input the enhanced feature matrix into the dynamic weighted multi-core support vector regression model for prediction and output preliminary prediction values. The model parameters and kernel weights of the dynamic weighted multi-core support vector regression model are obtained by optimizing the improved Super Flycatcher algorithm. The residual correction module is used to calculate the residual sequence of the preliminary prediction value, generate a correction increment based on the residual sequence using a temporal convolutional network, reconstruct the correction increment with the preliminary prediction value, and output the final wind power prediction value.

[0072] Specific limitations regarding the residual correction-based wind power prediction system can be found in the limitations of the residual correction-based wind power prediction method described above; the corresponding technical effects are equivalent and will not be repeated here. Each module in the aforementioned residual correction-based wind power prediction system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, allowing the processor to call and execute the corresponding operations of each module.

[0073] Example 3 Figure 2An internal structural diagram of a computer device is shown in one embodiment. This computer device may specifically be a terminal or a server. Figure 2 As shown, the computer device includes a processor, memory, network interface, display, camera, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a wind power prediction method based on residual correction. The display screen can be an LCD screen or an e-ink display screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0074] As will be understood by those skilled in the art, computer equipment Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0075] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0076] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0077] In summary, the wind power prediction method and system based on residual correction provided in this application use complete lumped empirical mode decomposition as a primary deconstruction method to extract several relatively independent intrinsic mode function components from the original non-stationary power signal. Using sample entropy as a complex quantification index, the intrinsic mode function components are further divided into noise, volatile, and trend terms, and a second deep stabilization based on variational mode decomposition is performed on the volatile term. Fuzzy information granulation technology is used to extract features from the quasi-stationary subsequence and trend term, outputting a multi-dimensional enhanced feature matrix containing low-order extreme values, mean, and high-order extreme values. An improved Super Falcon algorithm is used to optimize the penalty factor and kernel weights of the dynamic weighted multi-kernel support vector regression model across the entire domain, outputting preliminary prediction values ​​that take into account multi-scale signal characteristics. Finally, a temporal convolutional network is used to model the residual sequence of the preliminary prediction values, generating correction increments corresponding to the time step, and additively reconstructing the preliminary prediction values ​​to obtain the final wind power prediction value. The technical effects are concentrated in four aspects. First, the complexity-aware multi-level decomposition architecture significantly suppresses the mode aliasing phenomenon present in traditional empirical mode decomposition methods, and avoids over-processing of low-complexity components through on-demand secondary decomposition, achieving a balance between frequency band division and computational overhead. Second, the feature enhancement mechanism of fuzzy information granularization replaces point-by-point features in modeling with granular features, effectively suppressing micro-noise disturbances caused by wind field turbulence and improving the model's robustness to extreme weather. Third, the improved Super Flycatcher algorithm, which introduces the Lévy flight disturbance term, overcomes the shortcomings of traditional heuristic optimization methods that are prone to getting trapped in local optima in the later stages of iteration, providing the globally optimal penalty factor and kernel weight configuration for the dynamically weighted multi-kernel support vector regression model, enabling the model to simultaneously possess low-frequency trend fitting and high-frequency abrupt change response capabilities. Fourth, the closed-loop residual feedback architecture of the temporal convolutional network explicitly compensates for the prediction lag of the first-stage model at meteorological inflection points, ensuring that the overall prediction curve maintains a high coefficient of determination and low percentage error under normal fluctuation conditions, and significantly compresses extreme errors under meteorological abrupt change conditions.

[0078] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0079] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A wind power prediction method based on residual correction, characterized in that, include: Historical wind power sequences are obtained, and the historical wind power sequences are decomposed into several intrinsic mode function components using the fully lumped empirical mode decomposition technique. Calculate the sample entropy value of each intrinsic mode function component, and classify each intrinsic mode function component into noise term, volatile term and trend term based on the sample entropy value; The noise term is removed, and the volatile term is decomposed a second time using the variational mode decomposition algorithm to obtain a quasi-stationary subsequence; The quasi-stationary subsequence and the trend term are feature extracted using fuzzy information granulation technology to construct an enhanced feature matrix; The enhanced feature matrix is ​​input into the dynamic weighted multi-kernel support vector regression model for prediction, and the preliminary prediction value is output. The model parameters and kernel weights of the dynamic weighted multi-kernel support vector regression model are obtained by optimizing the improved Super Flycatcher algorithm. Calculate the residual sequence of the preliminary prediction value, use a temporal convolutional network to generate a correction increment based on the residual sequence, reconstruct the correction increment with the preliminary prediction value, and output the final wind power prediction value.

2. The wind power prediction method based on residual correction according to claim 1, characterized in that, The process of performing a primary decomposition of the historical wind power sequence using the fully lumped empirical mode decomposition technique yields several intrinsic mode function components, including: For multiple decomposition experiments, a standard normally distributed white noise sequence is added to the historical wind power sequence each time, and the sequence is decomposed to extract the corresponding temporary mode components. Calculate the ensemble average of the temporary modal components obtained from multiple decomposition experiments, and use the ensemble average as the eigenmode function component of the corresponding order.

3. The wind power prediction method based on residual correction according to claim 1, characterized in that, The step of classifying each intrinsic mode function component into noise, volatility, and trend terms based on the sample entropy value includes: A first entropy threshold and a second entropy threshold are preset, wherein the first entropy threshold is less than the second entropy threshold; Compare the sample entropy value with the first entropy threshold and the second entropy threshold: If the sample entropy value is less than the first entropy threshold, then the corresponding intrinsic mode function component is classified as the trend term; If the sample entropy value is greater than or equal to the first entropy threshold and less than the second entropy threshold, then the corresponding intrinsic mode function component is classified as the volatile term; If the sample entropy value is greater than or equal to the second entropy threshold, then the corresponding intrinsic mode function component is classified as the noise term.

4. The wind power prediction method based on residual correction according to claim 3, characterized in that, The step of using variational mode decomposition algorithm to perform secondary decomposition on the volatile term to obtain quasi-stationary subsequences includes: A variational optimization model for the volatile term is constructed with the goal of minimizing the sum of the estimated bandwidths of each mode. The volatile term is solved iteratively using the alternating direction multiplier method with the variational optimization model to separate several frequency band-limited modal components, which are then used as the quasi-stationary subsequence.

5. The wind power prediction method based on residual correction according to claim 1, characterized in that, The step of extracting features from the quasi-stationary subsequence and the trend term using fuzzy information granulation technology to construct an enhanced feature matrix includes: The quasi-stationary subsequence and the trend term are divided into sliding blocks according to a preset time window; Fuzzy membership functions are used to map data points within each time window, and the low-order extreme value, mean value, and high-order extreme value of the data within each time window are calculated and extracted. The low-order extreme values, mean values, and high-order extreme values ​​corresponding to each time window are combined and concatenated to construct the enhanced feature matrix.

6. The wind power prediction method based on residual correction according to claim 5, characterized in that, The combined kernel function in the dynamically weighted multi-kernel support vector regression model is a linear weighted sum of at least two basis kernel functions; The weight coefficients of each of the basis kernel functions are dynamic weighting coefficients, and the sum of the dynamic weighting coefficients of all basis kernel functions is equal to 1.

7. The wind power prediction method based on residual correction according to claim 1, characterized in that, The model parameters and kernel weights of the dynamically weighted multi-kernel support vector regression model are obtained through optimization using an improved Super Flycatcher algorithm, including: The improved super flycatcher algorithm updates the individual positions of the population using an individual position update equation that includes a Levy flight perturbation term during the iterative optimization process. The optimal individual position obtained after optimization convergence is analyzed and mapped to the penalty factor of the dynamically weighted multi-kernel support vector regression model and the dynamic weighting coefficient of each basis kernel function.

8. The wind power prediction method based on residual correction according to claim 7, characterized in that, The step of generating a correction increment using a temporal convolutional network based on the residual sequence includes: The residual sequence is input into the dilated convolutional layer of the temporal convolutional network to extract nonlinear error features from the time series; The nonlinear error features are mapped and output through the fully connected layer of the temporal convolutional network to obtain the correction increment for the corresponding time step.

9. A wind power prediction system based on residual correction, characterized in that, include: The primary decomposition module is used to obtain historical wind power sequences and perform primary decomposition on the historical wind power sequences using the fully lumped empirical mode decomposition technique to obtain several intrinsic mode function components. The complexity classification module is used to calculate the sample entropy value of each intrinsic mode function component and classify each intrinsic mode function component into noise, volatility and trend terms based on the sample entropy value. The secondary decomposition module is used to remove the noise terms and perform secondary decomposition on the volatile terms using the variational mode decomposition algorithm to obtain quasi-stationary subsequences. The feature enhancement module is used to extract features from the quasi-stationary subsequence and the trend term using fuzzy information granulation technology, and construct an enhanced feature matrix; The multi-core prediction module is used to input the enhanced feature matrix into the dynamic weighted multi-core support vector regression model for prediction and output preliminary prediction values. The model parameters and kernel weights of the dynamic weighted multi-core support vector regression model are obtained by optimizing the improved Super Flycatcher algorithm. The residual correction module is used to calculate the residual sequence of the preliminary prediction value, generate a correction increment based on the residual sequence using a temporal convolutional network, reconstruct the correction increment with the preliminary prediction value, and output the final wind power prediction value.

10. A computer device, characterized in that, include: Processor and computer-readable storage media; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the wind power prediction method based on residual correction as described in any one of claims 1 to 8.

Images