Short-term wind power prediction method and system based on variational mode decomposition and random forest

Abstract

The invention relates to a short-term wind power prediction method and system based on variational mode decomposition and random forest, and the method comprises the steps: decomposing wind power original data through a VMD algorithm, obtaining a plurality of intrinsic mode function components, and removing high-frequency noise through addition reconstruction; extracting statistical features of each IMF component, evaluating feature importance by using a random forest algorithm, and screening out redundant features; based on the screened features, constructing a deep combination model comprising CNN, BiGRU and MSA; an improved snow ablation optimization algorithm is introduced to automatically adjust key hyper-parameters of the CNN-BiGRU-MSA model; and through the trained combined model, predicting the wind power in a future period, and performing model performance evaluation. Compared with the prior art, the method has the advantages that the prediction precision and the model stability can be effectively improved, and the defects in the aspects of feature extraction, data noise reduction and model hyper-parameter optimization in the prior art are overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power prediction, in particular to a short-term wind power prediction method and system based on variational mode decomposition and random forest. BACKGROUND

[0002] Wind power generation, as a clean and renewable form of new energy, has achieved large-scale development worldwide. However, wind power is directly affected by wind speed fluctuations, and often changes dramatically in a short time scale. In particular, during periods of unstable weather and dramatic changes in wind speed, power fluctuations exhibit obvious nonlinear and non-stationary characteristics, posing serious challenges to real-time scheduling and safe and stable operation of the power grid.

[0003] Traditional wind power prediction methods rely on historical power data and numerical weather prediction information from wind farms, and combine statistical learning algorithms such as support vector machines, neural networks, and random forests to build power prediction models. However, these methods still have many limitations in terms of algorithm structure or data accuracy, especially when faced with complex and variable short-term wind speed disturbances, they cannot accurately depict their dynamic impact on wind power, resulting in large errors in short-term power prediction at the minute to hour level.

[0004] Wind speed data, as the main driving factor of wind power, has strong randomness, time-varying and nonlinearity. In a short time scale, wind speed signals often contain a large amount of high-frequency disturbances and noise. If these noises are not processed and directly input into the prediction model, they will significantly affect the stability and accuracy of the model. Therefore, how to extract key trends and remove disturbance components from the original wind speed or power sequence is one of the key problems in improving short-term wind power prediction.

[0005] In recent years, with the development of signal processing theory, variational mode decomposition (VMD) has been widely used in the preprocessing stage of wind power prediction due to its superior multi-scale decomposition ability. VMD can decompose complex wind power signals into a number of intrinsic mode functions (IMFs), separating different frequency band information, effectively suppressing high-frequency noise and retaining main trend components, providing clear and stable feature input for subsequent prediction modeling.

[0006] The main methods used in the prior art are: 1) Directly using a single machine learning model to model the original wind speed or power sequence, but without filtering the noise in the original sequence, resulting in significant disturbance of the model and unstable prediction results; 2) Introducing wavelet transform, empirical mode decomposition and other decomposition methods for signal preprocessing, which improves the model's ability to extract trend features, but has problems such as mode aliasing, serious end effect and insufficient decomposition accuracy; 3) Try to combine signal decomposition with machine learning model for combined modeling, such as EEMD-LSTM, WT-RF and other combined methods, but most of them lack systematic feature selection mechanism, have high model complexity and many redundant features, which can easily cause overfitting and reduce generalization performance.

[0007] Chinese patent application publication No. CN119561047A provides a wind power prediction method, system and equipment based on multi-model fusion, which solves the problems of volatility and intermittency in wind power prediction through a multi-model fusion method including data preprocessing, feature selection, VMD model optimization, CNN-ASSA-Informer model training and quantile regression, and improves prediction accuracy and reliability. However, using Pearson correlation for feature selection can easily ignore nonlinear feature correlation, and there is no clear process for identifying and removing high-frequency noise and boundary disturbances obtained by VMD decomposition, so that the subsequent model input is still affected by noise, affecting the stability of the prediction and the reliability of the confidence interval.

[0008] In summary, how to use an efficient signal decomposition method to obtain stable modal information, combine it with a machine learning model with strong robustness for modeling, and especially how to establish a fine modeling and prediction mechanism for non-stationary sequences, is the key to improving the accuracy of short-term wind power prediction, and is also a hot and difficult problem in current wind power prediction research. SUMMARY

[0009] The purpose of the present application is to overcome the defects of the prior art and provide a short-term wind power prediction method and system based on variational modal decomposition and random forest, to solve or partially solve the problem of unsatisfactory modeling ability and prediction accuracy for non-stationary wind power sequences of existing methods.

[0010] The purpose of the present application can be achieved by the following technical solutions: In one aspect of the present application, a short-term wind power prediction method based on variational modal decomposition and random forest is provided, comprising the following steps: Obtain the wind power time series, extract each modal component in the signal by variational modal decomposition iteration, identify and remove invalid modal components including high-frequency noise or boundary disturbance until the change threshold between modes meets the preset condition; Extract a plurality of time domain statistical indicators as candidate features for each modal component after removal, construct a random forest model based on multiple regression trees and train it, and select features associated with wind power; Use CNN to extract local spatial features, use BiGRU to extract time sequence features, and use MSA to calculate the correlation distribution between each time step and the global context, and construct a CNN-BiGRU-MSA model; Using features related to wind power as input, and combining optimal point set sampling and periodic oscillation mutations, the hyperparameters of the CNN-BiGRU-MSA model are searched and optimized to complete model training; Short-term wind power prediction is achieved based on the trained CNN-BiGRU-MSA model.

[0011] As a preferred technical solution, variational mode decomposition is achieved using the following formula: in, For modal number, These are the modal components derived from the signal. The center frequencies corresponding to each mode are: It is a pulse signal. This represents the total number of pre-decomposed items. Indicates time t Perform partial derivative operations. It is the imaginary unit.

[0012] As a preferred technical solution, for the random forest model, each regression tree extracts a subset of samples from the training data through bootstrap sampling during the training process, and performs node partitioning on the feature subset. Each regression tree uses internal nodes to represent the judgment conditions of candidate features, branches to represent the judgment results, and leaf nodes to represent the predicted result values ​​under the branch path.

[0013] As a preferred technical solution, the optimization objective of the random forest model is: in, Indicates the first Each sample number Indicates the first Each feature number Indicates the first 1 candidate split point This represents the left half of the region after the division. This represents the right half of the region after the division. , Representing regions and The optimal output value.

[0014] As a preferred technical solution, the candidate features include mean, variance, skewness, kurtosis, maximum value, and minimum value.

[0015] As a preferred technical solution, the CNN-BiGRU-MSA model is modeled as follows: CNN: In the formula, It is the first feature matrix output by the CNN. i Line number j The value of the column, ( ) is the activation function. It's the size of the sliding window. It is the convolution kernel n lines and m Column weights It is the input feature matrix n lines and m The value of the column, It is the kernel bias; BiGRU: In the formula, for t The door is constantly being updated. To reset the door, The current input is the feature output of the CNN after pooling and flattening. This is the current hidden state. Element-wise multiplication It is the Sigmoid activation function. To update the gate input weight matrix, To reset the gate input weight matrix, Input a weight matrix into the candidate hidden state. To update the hidden layer weight matrix of the gate, To reset the hidden layer weight matrix of the door, Let be the hidden layer weight matrix of the candidate hidden states. To update the gate bias term, To reset the door offset item, For candidate hidden state bias terms, MSA: The sequence features output by BiGRU are fed into attention heads of multiple scales for weighting, and the correlation distribution between each time step and the global context is calculated. The attention score matrix is ​​calculated as follows: In the formula, For BiGRU output t Time characteristics, For global context vectors, for the i-th time instant, t the attention weight at the i-th time instant, T denotes a preset time length, score() is an attention score function, for the i-th time instant, t the attention score at the i-th time instant, for the i-th time instant, k the attention score at the i-th time instant.

[0016] As a preferred technical solution, the process of searching and optimizing the hyperparameters of the CNN-BiGRU-MSA model includes the following steps: Initialize the model hyperparameter set, including the learning rate, the convolution kernel size, the number of BiGRU units, the number of attention heads, the training batch size, and the regularization coefficient, define the search range of each hyperparameter, and construct the solution space dimension; Randomly generate a matrix as a swarm, In each iteration, introduce a periodic oscillation mutation and a good point set strategy to update the individual position, guiding the individual to approach the current optimal solution and the overall centroid; For each individual corresponding to the hyperparameter combination, train the CNN-BiGRU-MSA model and calculate the prediction error index as the fitness function value, and record the current optimal solution; Set the maximum number of iterations or the fitness threshold as the termination condition, when the termination condition is met, output the hyperparameter combination corresponding to the current optimal individual, and obtain the trained CNN-BiGRU-MSA model.

[0017] As a preferred technical solution, the individual position is updated as follows: wherein, denotes t the updated position of the individual at the i-th time instant, i is the snowmelt rate, denotes a random number selected from [-1, 1], is the global historical optimal position, is the behavior weight or the moving step, is the element-wise multiplication, is the position of the individual at the i-th time instant, is the centroid position of all individuals at the i-th time instant. t i t

[0018] As a preferred technical solution, the process of identifying and removing invalid modal components including high-frequency noise or boundary disturbance includes the following steps: ​​​​All modal components are analyzed according to their spectral distribution and energy contribution, invalid modal components containing high-frequency noise or boundary disturbance are identified and removed, and the remaining low-frequency modal components are added and reconstructed to obtain the reconstructed wind power sequence after removing noise.

[0019] In another aspect of the present application, a short-term wind power prediction system based on variational mode decomposition and random forest is provided for implementing the aforementioned short-term wind power prediction method, comprising: A time series processing module is configured to obtain a wind power time series, extract each modal component in the signal through variational mode decomposition iteration, identify and remove invalid modal components including high-frequency noise or boundary disturbance until the change threshold between the modes meets the preset condition. A feature selection module is configured to extract a plurality of time domain statistical indicators as candidate features for each modal component after removal, construct a random forest model based on multiple regression trees and train the model to screen features associated with wind power. A model construction module is configured to extract local spatial features using CNN, extract time sequence features using BiGRU, and calculate the correlation distribution between each time step and the global context using MSA to construct a CNN-BiGRU-MSA model. A model training module is configured to use features associated with wind power as input, combine the sampling of good points and periodic oscillation mutations to search and optimize the hyperparameters of the CNN-BiGRU-MSA model, and complete model training. A prediction module is configured to implement short-term wind power prediction based on the trained CNN-BiGRU-MSA model.

[0020] Compared with the prior art, the present application has at least one of the following beneficial effects: (1) The present application uses variational mode decomposition (VMD) method to perform multi-scale signal decomposition on the original wind power sequence, decomposes the non-stationary signal into several intrinsic mode functions, effectively separates the trend component and high-frequency noise in the wind power, and improves the stationarity and predictability of the input data, significantly improving the response ability and robustness of the model to wind power mutations.

[0021] (2) The present application extracts time domain statistical features including mean, variance, skewness, kurtosis, maximum value and minimum value from each modal component obtained by VMD decomposition, and uses random forest algorithm to sort and filter the importance of the features. This mechanism can effectively remove redundant features, highlight key factors, reduce model dimension, and improve model training efficiency and prediction performance.

[0022] (3) The application constructs a hybrid deep neural network model of fusing convolutional neural network (CNN), bidirectional gated recurrent unit network (BiGRU) and multi-head self-attention mechanism (MSA), fully utilizes CNN to extract local spatial features, BiGRU to capture long-term time sequence dependency, and MSA to mine global correlation between features, so as to comprehensively improve the modeling capability of wind power time sequence data and realize accurate prediction of wind power.

[0023] (4) The application combines the snow melting optimization algorithm to perform self-adaptive search on the hyperparameters of the CNN-BiGRU-MSA hybrid model. The algorithm is based on the fusion of good point set and periodic oscillation mutation strategy, has strong global search capability and fast convergence capability, can effectively avoid falling into local optimum in the model training process, and improves the generalization capability and stability of the model.

[0024] (5) The application realizes systematic modeling of the nonlinear and non-stationary characteristics of wind power through a multi-stage integrated strategy of signal decomposition-feature screening-depth prediction-parameter optimization, ensures the prediction accuracy while enhancing the adaptability of the model to power fluctuations under different meteorological conditions, and is especially suitable for short-term power scheduling and safe operation scenarios under the condition of high proportion of wind power access to the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 It is a flowchart of the short-term wind power prediction method based on variational modal decomposition and random forest in the embodiment; Figure 2 It is a schematic diagram of the short-term wind power prediction process in the embodiment; Figure 3 It is a schematic diagram of data preprocessing architecture in the embodiment; Figure 4 It is a schematic diagram of the short-term wind power prediction system based on variational modal decomposition and random forest in the embodiment. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the application.

[0027] Embodiment 1 To solve the problems in the prior art, the embodiment provides a short-term wind power prediction method based on variational modal decomposition and random forest. The method combines signal processing technology and an integrated learning model to improve the modeling capability and prediction accuracy of non-stationary wind power sequences by decomposing, extracting features and modeling the historical data of wind power. A multi-scale signal processing method is introduced to denoise and reconstruct the power sequence, enhancing the adaptability of the model to wind speed mutations and power disturbances. An integrated learning model combined with a deep neural network is used to further improve the stability and generalization ability of the model. Compared with traditional methods, the method has significant advantages in handling short-term wind power prediction tasks and can effectively cope with wind speed fluctuations and nonlinear characteristics.

[0028] In the embodiment, the main principle of irradiance extraction based on classification prediction cloud image is: In the aspect of data preprocessing, there are usually a large amount of high-frequency noise and non-stationary disturbances in the original wind power sequence, which interfere with the prediction accuracy. To enhance the ability of the model to depict the trend of wind power changes, the original power signal is first decomposed into a plurality of intrinsic modal functions (IMF) with physical meaning by using the variational modal decomposition (VMD) method, effectively separating the high-frequency disturbance from the low-frequency trend information, and realizing signal denoising and multi-scale reconstruction. By observing the IMF component sequence, it can be found that there are obvious differences in local statistical characteristics between different modalities. In order to extract representative feature information, statistical feature parameters such as mean, variance, skewness, kurtosis, maximum and minimum are extracted on each modal component after decomposition, so as to construct a full-quantity feature vector as the input of the subsequent model.

[0029] In the aspect of feature selection, considering that the feature dimension obtained after multi-modal decomposition is high, the random forest (RF) algorithm is used to evaluate the importance of each statistical feature, and the key feature variables are selected accordingly, effectively reducing the feature dimension and improving the modeling efficiency and model robustness. At the same time, the random forest has strong non-linear modeling capability, which can preliminarily model the complex change relationship of wind power, and reduce the risk of overfitting while preserving the prediction basis.

[0030] In the aspect of prediction model, a CNN-BiGRU-MSA hybrid deep prediction model is constructed based on feature selection. The model combines convolutional neural network (CNN), bidirectional gated recurrent unit network (BiGRU) and multi-head self-attention mechanism (MSA), which are respectively used to extract local spatial features, capture long-term dependencies of time series and mine internal correlations among global features. The CNN structure can effectively capture mutations and boundary information in wind power sequences, the BiGRU can retain historical memory information in the sequence, and the MSA mechanism can enhance the expression of key features through attention allocation, thereby significantly improving the expression ability and prediction accuracy of the model.

[0031] To further improve the model performance, the embodiment combines the GVSAO (GVSAO) improved based on the best point set and the period mutation strategy, which is used for global optimization of the hyperparameters of the CNN-BiGRU-MSA model. The algorithm has strong exploration ability and convergence speed, can avoid the model from falling into local optimum, and effectively improves the final prediction effect. The combined prediction framework constructed by the method can comprehensively capture the trend change, noise disturbance and multi-scale characteristics in the wind power sequence, significantly improves the accuracy and reliability of short-term wind power prediction, and is especially suitable for high-frequency scheduling and power regulation demand in the high-penetration wind power grid-connected scene.

[0032] Next, the short-term photovoltaic power prediction method of the embodiment is introduced in detail, referring to Figure 1 and Figure 2 The method includes the following steps: S1, using VMD algorithm to decompose the original wind power data, obtaining a plurality of intrinsic mode function (IMF) components, and removing high-frequency noise by adding and reconstructing.

[0033] In the embodiment, in order to effectively alleviate the interference of a large amount of high-frequency noise and non-stationary disturbance existing in the original wind power data on subsequent modeling, the original wind power time series is decomposed by using the variational mode decomposition algorithm, and the original non-stationary signal is decomposed into a plurality of intrinsic mode function components with physical meaning.

[0034] Referring to Figure 3 S1 can specifically include the following sub-steps: S11, VMD calculates the frequency and amplitude of each mode by iteration algorithm, and gradually extracts different frequency components in the signal, and the specific model is: In the formula: is the mode number, is the mode component decomposed by the signal, is the center frequency corresponding to each mode, is the impulse signal, is the total number of pre-decomposition, which is set to 7 in the embodiment, indicates that the time partial derivative operation is performed, is the imaginary unit.

[0035] Select as the change threshold between modes, also known as the stopping criterion. If the change (such as frequency, amplitude, etc.) between modes is very small in a certain iteration, it means that the mode has tended to be stable, and the iteration can be stopped. When the change between all modes is less than the preset threshold after each iteration, VMD stops; S12. Analyze all IMF components according to their spectral distribution and energy contribution, identify and remove invalid mode components containing high-frequency noise or boundary disturbances. Sum and reconstruct the remaining mid-to-low frequency IMF components to obtain the noise-removed wind power reconstruction sequence, the mathematical expression of which is as follows: In the formula: This is the original signal sequence.

[0036] Through the above processing steps, the high-frequency random disturbances in the original wind power sequence are effectively weakened, while the main components reflecting the wind power operation trend and periodicity are retained, providing a stable and reliable input data foundation for subsequent feature extraction and predictive modeling.

[0037] S2. Extract the statistical features of each IMF component and use the random forest algorithm to evaluate the importance of the features and filter out redundant features.

[0038] In this embodiment, to further mine the wind power operation dynamic information contained in the IMF modal components after VMD decomposition and improve the input quality of subsequent prediction models, a combination of statistical feature extraction and random forest models is used to perform feature engineering processing on each modality. First, typical time-domain statistical indicators are extracted for each IMF component. Then, the relative importance of each feature is evaluated using a random forest model, and redundant information is removed. Finally, an optimized feature set for modeling is constructed. The specific implementation includes the following steps: S21. Regarding the regression nature of wind power prediction, this embodiment constructs a random forest model based on regression decision trees. This model consists of multiple regression trees. During training, each tree extracts a subset of samples from the training data using a bootstrap sampling method and partitions nodes on the feature subset. Each internal node in the regression tree represents a judgment condition for a feature attribute, each branch represents the judgment result, and each leaf node outputs the predicted value for that branch path. S22. To minimize regression error, Mean Squared Error (MSE) is used as the node splitting criterion during the training of each regression tree. Among all feature variables and their candidate splitting points, the feature and splitting point that minimizes the MSE loss are selected as the splitting criteria for the current node. The optimization objective function is expressed as follows: In the formula: —No. Each sample number —No. Each feature number —No. One dividing point, —The left half of the region after division — the right half region after division, , — the region and optimal output value; S23, time domain feature extraction is performed on the modal components obtained by decomposition, and 6 kinds of characteristic variables are obtained, including mean, variance, skewness, kurtosis, maximum value and minimum value. The above characteristic values are used as basic features to construct an initial feature vector set. Then input to the random forest model for training, and according to the feature importance ranking result generated by the model, the features with low contribution and strong redundancy are screened out, and only the variables with significant influence on the prediction performance are reserved as the final input, so as to improve the accuracy and efficiency of the subsequent power prediction model.

[0039] S3, based on the screened features, a deep combination model containing convolutional neural network (CNN), bidirectional gate recurrent unit (BiGRU) and multi-scale attention mechanism (MSA) is constructed.

[0040] After completing the statistical feature extraction and screening of the VMD decomposition modal, in order to further improve the modeling accuracy and time sequence sensitivity of the wind power prediction model, a deep combination prediction model integrating CNN, BiGRU and MSA modules is constructed. This model takes into account the spatial local feature extraction, time sequence information modeling and attention dynamic adjustment ability, and has strong feature expression ability and generalization ability.

[0041] Specifically includes the following sub-steps: S31, local spatial features of input feature data are extracted by using convolutional neural network (CNN), and the output calculation formula of convolutional layer is: In the formula: ( ) is an activation function, is the size of the sliding window, is the weight of n rows and m columns of the convolution kernel, is the value of n rows and m columns of the feature matrix of the input data, is the convolution kernel bias; S32, the time-dependent features of the sequence data are modeled by bidirectional gate recurrent unit (BiGRU), and the CNN output features are sent to the BiGRU network after pooling and flattening, and the forward and backward states in the time sequence data are jointly modeled. The specific formula is as follows: wherein: is an update gate, is a reset gate, is a current input, is a current hidden state, is an element-wise multiplication, is a sigmoid activation function, is an update gate input weight matrix, is a reset gate input weight matrix, is a candidate hidden state input weight matrix, is a hidden layer weight matrix for the update gate, is a hidden layer weight matrix for the reset gate, is a hidden layer weight matrix for the candidate hidden state, is an update gate bias term, is a reset gate bias term, is a candidate hidden state bias term.

[0042] The BiGRU adopts an update gate and a reset gate mechanism, effectively preserves important historical information, alleviates the gradient vanishing problem in traditional RNN networks, and improves the modeling ability of long-term dependency information; S33, to enhance the response ability of the model to key moments and core modal features, a multi-scale attention mechanism (MSA) module is introduced. This module sends the sequence features output by the BiGRU into multiple scale attention heads for weighting, and calculates the correlation distribution between each time step and the global context. Through attention weighted summation strategy, the response to high correlation features is strengthened, and invalid interference information is suppressed. The attention score matrix is calculated as follows: wherein: is a moment feature output by the BiGRU, is a global context vector, is the attention weight of the t moment, T denotes a preset time length, and score( ) is an attention score function, is the attention score of the t moment, is the attention score of the k moment; S34, input the high-dimensional feature vector processed by the CNN, BiGRU and MSA module into the fully connected layer, and output the final wind power prediction value. Optionally, the prediction error (such as MAE, RMSE, etc.) is calculated by comparing with the true value as the training optimization target, realizing the iterative update of the model in the training stage.

[0043] S4, in order to further improve the performance and generalization ability of the depth prediction model, an improved snowmelt optimization algorithm (GVSAO) is introduced to automatically search and optimize the key hyperparameters of the CNN-BiGRU-MSA model. The optimization strategy combines the good point set sampling and the periodic oscillation mutation mechanism, has strong global search ability and convergence efficiency, and is suitable for optimization tasks of complex high-dimensional model parameters. The specific implementation steps include: S41, set the model hyperparameter set to be optimized, including learning rate, convolution kernel size, BiGRU unit number, attention head number, training batch size, regularization coefficient, etc., define the search range of each hyperparameter, and construct the solution space dimension; S42, in SAO, the iteration process starts from a randomly generated swarm. The entire swarm is usually modeled as a matrix with rows and columns, and the specific formula is as follows: In the formula: represents the size of the swarm, represents the dimension of the solution space. Among them, and represent the lower and upper limits of the solution space, respectively. represents a randomly generated number in [0,1]; S43, in each iteration, the individual position is updated by introducing periodic oscillation mutation and good point set strategy, guiding the individual to approach the current optimal solution and the overall centroid. The position update formula is as follows: In the formula: is the snowmelt rate, is the global historical optimal position, is the behavior weight or moving step, is the element-wise multiplication, is t the position of the individual at time i , is the centroid position of all individuals at time t , represents a random number selected from [-1,1], this parameter helps the communication between individuals in this stage, with the help of crossover terms and , the individual is more likely to use the promising area according to the knowledge of the current best search agent and the group centroid position; S44, train the CNN-BiGRU-MSA model for each individual corresponding to the hyperparameter combination and calculate the prediction error index (such as MAE, RMSE, MAPE, etc.) as the fitness function value. Record the current optimal solution and use it as a guide for subsequent iterations; S45, set the maximum number of iterations or the fitness threshold as the termination condition, and output the hyperparameter combination corresponding to the current optimal individual when the termination condition is met.

[0044] S5, predict the future period of wind power by the trained combined model, and evaluate the model performance.

[0045] In this embodiment, after the training and hyperparameter optimization of the CNN-BiGRU-MSA combined model are completed, the model is applied to the wind power prediction task in the future period. Specifically, the test data after preprocessing and feature extraction is input, and the model automatically outputs the wind power prediction results of multiple time steps in the future, and the prediction performance of the model is evaluated by MAE, RMSE, MAPE.

[0046] Embodiment 2 Based on embodiment 1, this embodiment provides a short-term wind power prediction system based on variational modal decomposition and random forest, which is used to realize the short-term wind power prediction method of embodiment 1, and the system comprises: (1) Time series processing module, used to obtain wind power time series, extract each modal component in the signal through variational modal decomposition iteration, identify and eliminate invalid modal components including high-frequency noise or boundary disturbance, until the change threshold between modes meets the preset condition.

[0047] (2) Feature selection module, used to extract multiple time domain statistical indicators as candidate features for each modal component after elimination, construct a random forest model based on multiple regression trees and train it, and select features associated with wind power.

[0048] (3) Model construction module, used to extract local spatial features using CNN, extract time sequence features using BiGRU, calculate the correlation distribution between each time step and global context using MSA, and construct a CNN-BiGRU-MSA model.

[0049] (4) Model training module, used to search and optimize the hyperparameters of the CNN-BiGRU-MSA model with features associated with wind power as input, combined with good point set sampling and periodic oscillation mutation, and complete model training.

[0050] (5) Prediction module, used to realize short-term wind power prediction based on the trained CNN-BiGRU-MSA model.

[0051] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A short-term wind power prediction method based on variational mode decomposition and random forest, characterized in that, Includes the following steps: The wind power time series is obtained, and each mode component in the signal is extracted iteratively through variational mode decomposition. Invalid mode components, including high-frequency noise or boundary disturbances, are identified and removed until the change threshold between modes meets the preset conditions. For each modal component after removal, multiple time-domain statistical indicators are extracted as candidate features. A random forest model based on multiple regression trees is constructed and trained to filter out features associated with wind power. A CNN-BiGRU-MSA model is constructed by using CNN to extract local spatial features, BiGRU to extract temporal features, and MSA to calculate the correlation distribution between each time step and the global context. Using features related to wind power as input, and combining optimal point set sampling and periodic oscillation mutations, the hyperparameters of the CNN-BiGRU-MSA model are searched and optimized to complete model training; Short-term wind power prediction is achieved based on the trained CNN-BiGRU-MSA model.

2. The short-term wind power prediction method based on variational mode decomposition and random forest according to claim 1, characterized in that, Variational mode decomposition is achieved using the following formula: in, For modal number, These are the modal components derived from the signal. The center frequencies corresponding to each mode are: It is a pulse signal. This represents the total number of pre-decomposed items. Indicates time t Perform partial derivative operations. It is the imaginary unit.

3. The short-term wind power prediction method based on variational mode decomposition and random forest according to claim 1, characterized in that, For the random forest model, each regression tree extracts a subset of samples from the training data during the training process using bootstrap sampling, and performs node partitioning on the feature subset. Each regression tree uses internal nodes to represent the judgment conditions of candidate features, branches to represent the judgment results, and leaf nodes to represent the predicted result values ​​under the branch path.

4. The short-term wind power prediction method based on variational mode decomposition and random forest according to claim 1, characterized in that, The optimization objective of the random forest model is: in, Indicates the first Each sample number Indicates the first Each feature number Indicates the first 1 candidate split point This represents the left half of the region after the division. This represents the right half of the region after the division. , Representing regions and The optimal output value.

5. The short-term wind power prediction method based on variational mode decomposition and random forest according to claim 1, characterized in that, The candidate features include mean, variance, skewness, kurtosis, maximum value, and minimum value.

6. The short-term wind power prediction method based on variational mode decomposition and random forest according to claim 1, characterized in that, The CNN-BiGRU-MSA model is modeled as follows: CNN: In the formula, It is the first feature matrix output by the CNN. i Line 1 j The value of the column, ( ) is the activation function. It's the size of the sliding window. It is the convolution kernel n lines and m Column weights It is the input feature matrix n lines and m The value of the column, It is the kernel bias; BiGRU: In the formula, for t The door is constantly being updated. To reset the door, The current input is the feature output of the CNN after pooling and flattening. This is the current hidden state. Element-wise multiplication It is the Sigmoid activation function. To update the gate input weight matrix, To reset the gate input weight matrix, Input a weight matrix into the candidate hidden state. To update the hidden layer weight matrix of the gate, To reset the hidden layer weight matrix of the door, Let be the hidden layer weight matrix of the candidate hidden states. To update the gate bias term, To reset the door offset item, For candidate hidden state bias terms, MSA: The sequence features output by BiGRU are fed into attention heads of multiple scales for weighting, and the correlation distribution between each time step and the global context is calculated. The attention score matrix is ​​calculated as follows: In the formula, For BiGRU output t Time characteristics, For global context vectors, For the first t Attention weight at any moment T This indicates the preset time length, and score() is the attention scoring function. For the first t Attention score at any moment For the first k Attention score at any given moment.

7. The short-term wind power prediction method based on variational mode decomposition and random forest according to claim 1, characterized in that, The process of searching and optimizing the hyperparameters of the CNN-BiGRU-MSA model includes the following steps: Initialize the model hyperparameter set, including learning rate, kernel size, number of BiGRU units, number of attention heads, training batch size, and regularization coefficient; define the search range for each hyperparameter; and construct the solution space dimension. Randomly generated matrices are used as swarms. In each iteration, a periodic oscillation mutation and a best point set strategy are introduced to update the individual position, guiding the individual to approach the current optimal solution and the centroid of the whole; For each individual's corresponding hyperparameter combination, train the CNN-BiGRU-MSA model and calculate the prediction error index as the fitness function value, and record the current optimal solution. Set a maximum number of iterations or a fitness threshold as a termination condition. When the termination condition is met, output the hyperparameter combination corresponding to the current best individual to obtain the trained CNN-BiGRU-MSA model.

8. The short-term wind power prediction method based on variational mode decomposition and random forest according to claim 7, characterized in that, Update the individual position using the following formula: in, express t Individual at +1 time i The updated location It is the snow melting rate. This represents a random number selected from [-1, 1]. This is the best position in the global history. For behavior weights or movement steps, For element-wise multiplication, for t individual moment i Location, for t The position of the center of mass of all individuals at any given moment.

9. The short-term wind power prediction method based on variational mode decomposition and random forest according to claim 1, characterized in that, The process of identifying and eliminating invalid modal components, including high-frequency noise or boundary perturbations, includes the following steps: All modal components are analyzed according to their spectral distribution and energy contribution. Invalid modal components containing high-frequency noise or boundary disturbances are identified and removed. The remaining mid- and low-frequency modal components are summed and reconstructed to obtain the wind power reconstructed sequence after noise removal.

10. A short-term wind power prediction system based on variational mode decomposition and random forest, characterized in that, For implementing the short-term wind power forecasting method as described in any one of claims 1-9, the system comprises: The time series processing module is used to acquire wind power time series, extract each mode component in the signal through variational mode decomposition iteration, identify and remove invalid mode components including high-frequency noise or boundary disturbances, until the change threshold between modes meets the preset conditions. The feature selection module is used to extract multiple time-domain statistical indicators as candidate features for each modal component after elimination, construct a random forest model based on multiple regression trees and train it, and filter out features related to wind power. The model building module is used to extract local spatial features using CNN, extract temporal features using BiGRU, and calculate the correlation distribution between each time step and the global context using MSA to build a CNN-BiGRU-MSA model. The model training module is used to search and optimize the hyperparameters of the CNN-BiGRU-MSA model by taking features related to wind power as input, combining optimal point set sampling and periodic oscillation mutations, and completing the model training. The prediction module is used to predict short-term wind power based on the trained CNN-BiGRU-MSA model.

Images