CN-121983970-A - Wind power prediction method and system based on frequency domain self-adaptive super-parameter optimization

CN121983970ACN 121983970 ACN121983970 ACN 121983970ACN-121983970-A

Abstract

The invention discloses a wind power prediction method and a wind power prediction system based on frequency domain self-adaptive super-parameter optimization, wherein the method comprises the steps of firstly collecting historical wind power and meteorological data; the method comprises the steps of providing a brand new frequency domain coverage overlap coefficient as a fitness function, combining a gray wolf optimization algorithm to adaptively optimize super parameters of variation modal decomposition, decomposing a power sequence by utilizing the optimized parameters, calculating sample entropy of each component and reconstructing the sample entropy into three types of random, fluctuation and trend according to entropy values to reduce complexity, screening weather features which are strongly related for each component based on pearson correlation coefficients, respectively inputting a prediction model to predict, and finally superposing predicted values of each component to obtain a final power prediction result. The method solves the problem of manual setting of the VMD super parameters, realizes balance of decomposition quality and prediction efficiency, and remarkably improves accuracy and practicability of ultra-short-term wind power prediction.

Inventors

HE HAO
ZHAO WEIZHE
WU KANG
HUANG XU
HU CHENG
GUO LIANG

Assignees

国网江西省电力有限公司电力科学研究院

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. The wind power prediction method based on frequency domain self-adaptive super-parameter optimization is characterized by comprising the following steps of: acquiring historical wind power data of a target area and historical meteorological data associated with the historical wind power data, and dividing the historical wind power data into a training set and a testing set according to a preset proportion; Defining a frequency domain coverage overlap coefficient based on a frequency domain, and using a gray wolf optimization algorithm to adaptively optimize the total number of modes of the variation modal decomposition and a penalty factor by taking the frequency domain coverage overlap coefficient as an optimizing target to obtain an optimal super-parameter combination; Decomposing the wind power sequences in the training set and the test set by utilizing the variation modal decomposition model configured by the optimal super-parameter combination to obtain a plurality of eigen modal function components; Calculating sample entropy values of the plurality of eigenmode function components, and reconstructing all eigenmode function components into random components, fluctuation components and trend components according to the size range of the sample entropy values; Based on the pearson correlation coefficient, respectively screening a first feature subset corresponding to the random component, a second feature subset corresponding to the fluctuation component and a third feature subset corresponding to the trend component from the historical meteorological data; inputting the first feature subset, the second feature subset and the third feature subset into a preset prediction model, and outputting the prediction model to obtain a corresponding random component predicted value, a fluctuation component predicted value and a trend component predicted value; And superposing the random component predicted value, the fluctuation component predicted value and the trend component predicted value to obtain a final wind power predicted result.
2. The wind power prediction method based on frequency domain adaptive super-parameter optimization according to claim 1, wherein the calculation formula of the frequency domain coverage overlap coefficient is: , In the formula, For the maximum value of the frequency interval after the fast fourier transform, For the total number of modes of the VMD decomposition, For the magnitude of the kth order IMF component at frequency ω, For the magnitude of the k +1 order IMF component at frequency ω, Is the magnitude of the wind power at frequency ω.
3. The wind power prediction method based on frequency domain adaptive super-parameter optimization according to claim 1, wherein the obtaining the optimal super-parameter combination by using a gray wolf optimization algorithm and taking the frequency domain coverage overlap coefficient as an optimization target to adaptively optimize the total number of modes of the variation modal decomposition and a penalty factor comprises: initializing a gray wolf population, and setting the optimizing interval of the total number of modes and the punishment factors; in each iteration, calculating a frequency domain coverage overlap coefficient under the super-parameter combination corresponding to the current gray wolf position; Updating the gray wolf population according to a position updating mechanism of a gray wolf optimization algorithm, and recording a historical optimal solution; and outputting a super-parameter combination with the minimum frequency domain coverage overlap coefficient value as the optimal super-parameter combination when the preset maximum iteration number is reached.
4. The wind power prediction method based on frequency domain adaptive super-parameter optimization according to claim 1, wherein the decomposing the wind power sequences in the training set and the test set by using the variation modal decomposition model configured by the optimal super-parameter combination to obtain a plurality of eigenvalue function components comprises: Constructing a constraint variation problem, wherein the objective of the constraint variation problem is to minimize the sum of the estimated bandwidths of the components of each eigenmode function, the constraint condition of the constraint variation problem is that the sum of all the components is equal to an original power sequence, and the expression is as follows: , In the formula, For the set of decomposed IMFs, To resolve the set of center frequencies of the IMF, In order to resolve the IMF number of the wafer, For time of Is used for the partial derivative operation of (a), In order to achieve a peripheral rate of the material, In units of imaginary numbers, Is the first The IMF of the order of the IMF, Is the first The center frequency of the IMF of order, The method is an original wind power sequence; Introducing a secondary penalty factor and a Lagrangian multiplier, and converting the constraint variation problem into an augmented Lagrangian function, wherein the expression is as follows: , In the formula, In order to be a lagrange multiplier, In order to penalize the coefficients, As a lagrangian multiplier function in the time domain, Is a dirac function; solving the augmented Lagrangian function by adopting an alternate direction multiplier method, and alternately updating the center frequency and the amplitude of each eigenmode function component and the Lagrangian multiplier in a frequency domain until convergence conditions are met to obtain a plurality of decomposed eigenmode function components, wherein the expression is as follows: , In the formula, Is the first The first iteration The fourier transform of the IMF-order, Is the fourier transform of the original wind power sequence, Is the first The first iteration The fourier transform of the IMF-order, Is the first The first iteration The fourier transform of the IMF-order, Is the first Fourier transform of lagrangian multipliers at a number of iterations, In order to penalize the coefficients, Is the first The first iteration The center frequency of the IMF of order, Is the first Fourier transform of lagrangian multipliers at iteration, Is the first Fourier transform of lagrangian multipliers at a number of iterations, Is noise tolerance.
5. The wind power prediction method based on frequency domain adaptive super-parameter optimization according to claim 1, wherein reconstructing all eigenvalue function components into random components, fluctuation components and trend components according to the magnitude range of sample entropy values comprises: Setting a first entropy threshold and a second entropy threshold, wherein the first entropy threshold is smaller than the second entropy threshold; classifying eigenmode function components with sample entropy values greater than the second entropy threshold as random components; Classifying eigenmode function components with sample entropy values between the first entropy threshold value and the second entropy threshold value as fluctuation components; the eigenmode function components with sample entropy values smaller than the first entropy threshold are classified as trend components.
6. The wind power prediction method based on frequency domain adaptive hyper-parameter optimization of claim 1, wherein the preset prediction model comprises a long-short-term memory network model for predicting the trend component, a two-way long-short-term memory network model for predicting the fluctuation component, and a two-way long-short-term memory network model for predicting the random component and introducing attention mechanism.
7. The wind power prediction method based on frequency domain adaptive super-parameter optimization according to claim 1, wherein the filtering the first feature subset corresponding to the random component, the second feature subset corresponding to the fluctuation component, and the third feature subset corresponding to the trend component from the historical meteorological data based on pearson correlation coefficients comprises: Calculating pearson correlation coefficients of each meteorological feature in the historical meteorological data and a current component, wherein the current component is the random component, the fluctuation component or the trend component; And selecting all meteorological features with absolute values of pearson correlation coefficients larger than a preset threshold value to form a feature subset corresponding to the current component.
8. Wind power prediction system based on frequency domain self-adaptive super-parameter optimization is characterized by comprising: The acquisition module is configured to acquire historical wind power data of a target area and historical meteorological data associated with the historical wind power data, and divide the historical wind power data into a training set and a testing set according to a preset proportion; The optimizing module is configured to define a frequency domain coverage overlapping coefficient based on a frequency domain, and to adaptively optimize the total number of modes of the variation modal decomposition and a penalty factor by taking the frequency domain coverage overlapping coefficient as an optimizing target by utilizing a gray wolf optimizing algorithm to obtain an optimal super-parameter combination; The decomposition module is configured to decompose wind power sequences in a training set and a test set respectively by utilizing the variation modal decomposition model configured by the optimal super-parameter combination to obtain a plurality of intrinsic modal function components; The reconstruction module is configured to calculate sample entropy values of the plurality of eigenmode function components and reconstruct all eigenmode function components into random components, fluctuation components and trend components according to the size range of the sample entropy values; A screening module configured to screen a first feature subset corresponding to the random component, a second feature subset corresponding to the fluctuating component, and a third feature subset corresponding to the trend component from the historical meteorological data, respectively, based on pearson correlation coefficients; the prediction module is configured to input the first feature subset, the second feature subset and the third feature subset into a preset prediction model, and the prediction model outputs a random component predicted value, a fluctuation component predicted value and a trend component predicted value which are corresponding to each other; and the superposition module is configured to superpose the random component predicted value, the fluctuation component predicted value and the trend component predicted value to obtain a final wind power predicted result.
9. An electronic device comprising at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method of any of claims 1 to 7.

Description

Wind power prediction method and system based on frequency domain self-adaptive super-parameter optimization Technical Field The invention belongs to the technical field of electric power automation, and particularly relates to a wind power prediction method and system based on frequency domain self-adaptive super-parameter optimization. Background At present, wind power prediction methods can be mainly divided into a physical method, a statistical method and an artificial intelligence method. The artificial intelligence method based on deep learning, such as long-term and short-term memory network and its variants, is widely used because of its strong nonlinear time sequence modeling capability. However, the original wind power sequence has high non-stationarity and complexity, the fluctuation characteristics of multiple time scales of the original wind power sequence are difficult to be fully captured by a single deep learning model, and the prediction accuracy has a bottleneck. In order to improve the processing power for non-stationary sequences, the strategy of "decomposition-prediction-reconstruction" is widely adopted. The strategy firstly utilizes a signal decomposition technology to decompose an original power sequence into a plurality of relatively stable subsequences (intrinsic mode functions, IMFs), then respectively models and predicts each subsequence, and finally aggregates a prediction result. The variational modal decomposition is used as a non-recursive self-adaptive signal decomposition method, and by constructing and solving the constraint variational problem, the signal can be adaptively decomposed into modal components with specific sparsity in a frequency domain, so that compared with methods such as empirical modal decomposition, the modal aliasing problem is effectively avoided, and the decomposition stability is stronger. However, the decomposition effect of the VMD is severely dependent on two key superparameters, the modal number K and the penalty factor α. K determines the granularity of decomposition, too small can lead to under decomposition, too large can introduce noise and increase calculation load, and alpha controls the bandwidth of each modal component and directly influences the frequency domain concentration and mutual independence of the components. In practical application, the parameters are usually manually preset by relying on expert experience or grid search, and lack of self-adaptive capability, so that the optimality of a decomposition result in a frequency domain is difficult to ensure, and the upper performance limit of a subsequent prediction model is restricted. In addition, if an independent complex prediction model is built for each decomposed IMF component, the calculation cost will rise sharply, and it is difficult to meet the requirement on efficiency in engineering application. Therefore, how to adaptively determine the optimal super parameters of the VMD, and reduce the complexity of the model on the premise of ensuring the prediction accuracy, becomes a key technical problem to be solved in order to improve the ultra-short-term wind power prediction performance. Disclosure of Invention The invention aims to overcome the defects of the prior art, provides a wind power prediction method and a wind power prediction system based on frequency domain self-adaptive super-parameter optimization, according to the invention, a brand new frequency domain evaluation index is designed to adaptively optimize VMD parameters, intelligent reconstruction and differential modeling are performed based on component complexity, and the calculation complexity is effectively controlled while the prediction accuracy is improved. In a first aspect, the present invention provides a wind power prediction method based on frequency domain adaptive super parameter optimization, including: acquiring historical wind power data of a target area and historical meteorological data associated with the historical wind power data, and dividing the historical wind power data into a training set and a testing set according to a preset proportion; Defining a frequency domain coverage overlap coefficient based on a frequency domain, and using a gray wolf optimization algorithm to adaptively optimize the total number of modes of the variation modal decomposition and a penalty factor by taking the frequency domain coverage overlap coefficient as an optimizing target to obtain an optimal super-parameter combination; Decomposing the wind power sequences in the training set and the test set by utilizing the variation modal decomposition model configured by the optimal super-parameter combination to obtain a plurality of eigen modal function components; Calculating sample entropy values of the plurality of eigenmode function components, and reconstructing all eigenmode function components into random components, fluctuation components and trend components according to the size range of the sample