Wind power combination prediction method and system based on successive variational mode decomposition
Through successive variational modal decomposition and fuzzy entropy algorithm, wind power data is divided into high-frequency and low-frequency components, and prediction is performed using CLformer and AESELM models respectively. Finally, weighted summing is performed through entropy weight method, which solves the problem of limited accuracy of the existing wind power power prediction model, and achieves a more efficient and stable prediction effect.
Patent Information
- Application Number
- CN202411590785.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-08
- Publication Date
- 2025-03-28
AI Technical Summary
The existing wind power power prediction model has limited prediction accuracy, making it difficult to effectively capture different frequency components in wind power power data, resulting in unstable prediction results.
Sequential variational modal decomposition (SVMD) is used to decompose wind power power data into multiple components, combine fuzzy entropy algorithm (FE) to divide high-frequency and low-frequency components, and use CLformer and AESELM models for prediction respectively, and finally weighted summing is performed through entropy weight method to obtain the final prediction result.
Through the method of sequential decomposition and multi-model prediction, the accuracy and stability of wind power prediction are significantly improved, and different frequency components in the data can be better captured, and the prediction ability of wind farm operation is enhanced.
Smart Images
Figure CN119721742A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of wind power prediction, and relates to a wind power combination prediction method and system based on successive variational modal decomposition. Background Art
[0002] As a clean energy technology, wind power has significant professional advantages in the energy field. However, wind power is very susceptible to natural factors, and its power generation is unstable, fluctuating and intermittent. This will affect the power grid's arrangement of power dispatch and cause certain economic losses. Accurate wind power forecasting can help power grid operators and wind farm managers to reasonably arrange power dispatch and improve the stability and reliability of the power grid. In addition, wind power forecasting can also help plan and optimize the operation strategy of wind farms, improve wind power generation efficiency, and reduce the operating costs of wind farms. Therefore, in order to maintain the stable and efficient operation of the power grid system, wind power forecasting is extremely necessary.
[0003] At present, wind power prediction models can be divided into three categories: physical models, statistical models, and artificial intelligence models. The physical model considers numerical weather prediction and further relies on the dynamic model of atmospheric flow in order to obtain the corresponding wind speed estimate at the hub height of the wind turbine. Then, the idealized wind speed power curve is used to obtain the corresponding estimate of wind power output. Finally, the model output is used to correct the error of the residual model. Unlike physical models, statistical models aim to find the relationship between power measurements and some explanatory variables. Artificial intelligence models have become the mainstream model for wind power prediction due to their advantages such as fast calculation speed, high precision, and wide range of applications. At present, common artificial intelligence prediction models include long short-term memory recursive neural networks, feedback neural networks, recurrent neural networks, extreme learning machines, etc. The limitations of a single prediction model lead to limited prediction accuracy. Therefore, the above defects are problems that need to be solved. Summary of the invention
[0004] Purpose of the invention: The purpose of the present invention is to provide a method and system for wind power combination prediction based on successive variational mode decomposition in response to the problems in the background technology.
[0005] Technical solution: A wind power combination prediction method based on successive variational modal decomposition of the present invention comprises the following steps:
[0006] Step 1: Obtain historical power data of the wind farm and perform data cleaning;
[0007] Step 2: Decompose the historical wind power data into k components using SVMD;
[0008] Step 3: Use the fuzzy entropy algorithm FE to perform a complexity analysis on the wind power subsequence components decomposed by SVMD and divide them into high-frequency components and low-frequency components, and divide them into a training set and a test set;
[0009] Step 4: Establish a CLformer model, and send the high-frequency component training set obtained in step (3) into the CLformer model for model training; the CLformer model has high accuracy in predicting fast-changing high-frequency time series;
[0010] Step 5: Establish the AESELM model, and send the low-frequency component training set in step (3) into the AESELM model for model training; the AESELM model can more accurately capture the characteristics and trend information of the low-frequency components;
[0011] Step 6: The quadratic interpolation optimization algorithm QIO is improved and optimized by using Cubic chaotic map initialization and Gaussian random walk strategy to obtain the IQIO algorithm; the hyperparameters of the CLformer model are optimized by using the quadratic interpolation optimization algorithm IQIO;
[0012] Step 7: Send the high-frequency component test set data of step (3) into the trained and optimized CLformer model, and send the low-frequency component test set into the AESELM model to predict the wind power of the wind farm respectively, and perform weighted summation on the measurement results based on the entropy weight method to obtain the final prediction result.
[0013] Furthermore, the step 2 decomposes the input signal by SVMD, and the steps are as follows:
[0014] Step 2.1, SVMD can decompose complex data into multiple components without setting the number of modes, and because SVMD has strong anti-noise ability, it is particularly suitable for processing nonlinear and non-stationary signals. The original wind power data is decomposed into two parts: k components and residual signals. The decomposition formula is as follows:
[0015] f(t)=u k (t)+f r (t) (1)
[0016] Among them, the residual signal f r (t) includes the unprocessed part f u (t) and the sum of the previously obtained components
[0017] Step 2.2, define the final decomposition formula:
[0018]
[0019] Where k represents the total number of components of the wind power data subsequence, u k (t) is the kth wind power data component, is the sum of the components of the wind power data subsequence obtained previously, f u (t) is the wind power data that has not been processed.
[0020] Furthermore, the implementation process of step 3 is as follows:
[0021] Step 3.1: For the time series X (t) , the fuzzy entropy FE calculation method is as follows:
[0022]
[0023]
[0024]
[0025] Where m is the embedding dimension, is the similarity limit; X i , X j For the time series X (t) subsequence in For X i , X j The values between them are the maximum values of their respective endpoint differences; r is the similarity; n is the gradient; is the constructor; S is X (t) length;
[0026] Step 3.2: For the components f1…f of SVMD decomposition k , where k is the total number of components, calculate its fuzzy entropy value, then compare the fuzzy entropy differences between adjacent components, and select the two components with the largest fuzzy entropy difference f i and f i+1 , the components f1~f i Defined as high-frequency component, the component f i+1 ~f k Defined as the low frequency component.
[0027] Furthermore, in step 4, the high-frequency wind power subsequence output from the fuzzy entropy FE is input into the CLformer for prediction, and the implementation process is as follows:
[0028] The CLformer model is a lightweight model based on the Transformer architecture. The LGAC mechanism is proposed in the CLformer model. This mechanism calculates autocorrelations within equal-length sequence segments, enhancing the model's ability to extract local temporal dynamics. By setting different sequence grouping lengths, LGAC can communicate between non-overlapping sequence segments and capture multi-scale temporal patterns.
[0029] Step 4.1: Convert the high-frequency time series of wind power X en ∈R L×m Input into the encoder and then transformed into X through the embedding layer en ∈R L×D ; Among them, X en It represents the matrix of high-frequency time series; R L×m represents a set of high-frequency real numbers with L rows and m columns, R L×D It represents a high-frequency real number set encoder with L rows and D columns;
[0030] Step 4.2: Encoder input is X en ∈R L×m , the encoder output is represents a real data set associated with high frequencies at N time steps, and D is the dimension of the hidden state; the encoder calculation process is as follows:
[0031] X t =AvgPool(Padding(X))
[0032] X s =XX t
[0033] X t =SeriesDecomp(X)
[0034]
[0035]
[0036]
[0037] in They represent the seasonal components obtained from the i-th time series decomposition component of the l-th layer, respectively, where l∈{1,…,}, TCB is the time convolution function, FeedForward is the feedforward neural network function, SeriesDecomp is the sequence decomposition function, AvgPool is the average function, Padding is the padding function, and X t is the result of decomposing the filled X sequence, X s For X and X t The difference between
[0038] Step 4.3: and In the input decoder, O is the prediction range; It represents the dataset of the initial time step when decoding the seasonal component in wind power forecasting; It represents the data set of the initial time step when decoding the trend period component in wind power forecasting; and Medium length The encoder input The seasonal and trend components are decomposed from the second half of ; the decoder calculation process is as follows:
[0039]
[0040] in represents the seasonal and trend components of the i-th time series decomposition component of the l-th layer; LGAC is the local group autocorrelation function; where l∈{1,...,M}, M is the number of decoder layers with the same decoder, ω l ∈R D ×n Represents the forecast trend component The sum of s Indicates that The dimension is converted from D to the projection matrix of the target dimension n; Sum and In projection Afterwards, the results are output as a time series of CLformer predictions.
[0041] Furthermore, in step 5, the low-frequency wind power subsequence output from the fuzzy entropy FE is input into the AESELM model for prediction. The prediction process is as follows:
[0042] The AESELM model inherits the fast learning characteristics of ELM, but introduces an adaptive mechanism based on ELM, which means that the model can dynamically adjust its parameters according to the characteristics of the data, which can effectively improve the model's search ability and generalization ability, and effectively improve the prediction accuracy.
[0043] Step 5.1: Map the low-frequency wind power data from the input space to the AESELM feature space. The output of AESELM is:
[0044]
[0045] Among them, L is the number of hidden layer nodes, β i is the weight of the hidden layer and the output layer, h i (x) is the activation function; H is the hidden layer output matrix, β is the output weight matrix;
[0046] Step 5.2: Define the activation function as g(x) through the recorded training set, select the number of nodes L in the hidden layer, randomly generate input weights ωi and hidden layer bias i, calculate the hidden layer output matrix H, and obtain the wind power prediction result χ:
[0047]
[0048] Among them, H + is the generalized inverse matrix of H, N is the number of input wind power sequences, and T is the desired target matrix.
[0049] Furthermore, the step 6 is implemented as follows:
[0050] Step 6.1, set the population size and number of iterations of the IQIO algorithm as well as the upper and lower limits of the search space;
[0051] Step 6.2: Improve the population initialization of the original QIO algorithm and use the good point set strategy to initialize the population of the QIO algorithm;
[0052]
[0053] p n (y) = {(L1y1,L2y2,...L n y n )},y=1,2,3...n
[0054]
[0055] Among them, a y is the lower dimension of the current dimension, b j is the upper dimension of the current dimension, γ i is the i-th individual, L j For the best point, p n (y) is the set of good points, Updated population.
[0056] Step 6.3, define the objective function as the mean absolute error (MAE) between the predicted value and the actual value of the wind power forecast event, and calculate the fitness value of the population through the objective function;
[0057] Step 6.4, individual update; search for the optimal hyperparameters of the CLformer model through different update methods. The hyperparameters include the number of potential factors in matrix decomposition and the number of hidden layers of the deep neural network, so that the average value of the absolute error between the predicted value and the actual value of the wind power forecast event is reduced;
[0058] Step 6.5: Continuously update the current solution through algorithm iteration, and simultaneously update the global optimal solution;
[0059] Step 6.6: Determine whether the algorithm termination condition is reached through the given maximum number of iterations, and finally output the optimal hyperparameters of the CLformer model within the maximum number of iterations.
[0060] Furthermore, the individual positions in step 6.4 are updated using the exploration strategy:
[0061]
[0062] ξ1=γ i+1 (m),γ rand1 (m),γ rand2 (m)
[0063]
[0064] w1=3n1b
[0065]
[0066]
[0067]
[0068] Where * represents the GIQ function of the number, γ rand represents individuals at different positions randomly selected from the group, ξ is the constructor, v i (m+1) indicates that at the m+1th iteration, individual γ i , w represents the exploration weight, n follows the standard normal distribution, b is the adaptive parameter, m is the number of iterations, M is the maximum number of iterations, e is the control parameter of b, p is a random number (0, 1), Represents the round function;
[0069] The valid value of GIQ function is: f(x) = x 2 -10cos(2πx)+10x∈[0.6,2].
[0070] Furthermore, in step 6.4, the individual positions are updated using the development strategy:
[0071]
[0072] ξ2=γ best (m),γ rand1 (m),γ rand2 (m)
[0073]
[0074]
[0075]
[0076] Introduce Gaussian random walk to replace the individual position update of the original QIO algorithm:
[0077]
[0078] σ=γ(m)-γ rand (m) (13)
[0079] Among them, pD represents a random integer in [1, d], Ub pD ,Lb pD are the upper and lower limits of the pDth dimension, ε i+1 has a mean of 0 and a variance of σ 2 A normally distributed random variable.
[0080] Furthermore, the implementation process of step 7 is as follows:
[0081] The high-frequency component test set is sent to the trained and optimized CLformer model for prediction, and the low-frequency component test set is sent to the trained AESELM model for prediction to obtain the prediction results of each component; at the same time, the component results are weighted and summed based on the entropy weight method. The entropy weight method first standardizes the data:
[0082]
[0083] Among them l ij is the normalized value of the i-th component on the j-th index, x ij is the value of the i-th component at the j-th index, x max ( j ) is the maximum value on the jth index, x min ( j ) is the minimum value on the jth index;
[0084] After standardizing each component, calculate the weight of each component under each indicator:
[0085]
[0086] where p ij It represents the weight of the i-th component under the j-th index;
[0087] Then calculate the information entropy and entropy weight of each indicator:
[0088]
[0089]
[0090] where e j 、w j are the information entropy and weight of the jth indicator, n and m are the number of components and the number of indicators, respectively. is a constant used to normalize entropy values;
[0091] Finally, the calculated entropy value is used to calculate the comprehensive score of each component:
[0092]
[0093] The final weight of each component is obtained by calculating the comprehensive score of each component, and the final wind power prediction result is obtained according to the weight.
[0094] A wind power combination prediction system based on successive variational mode decomposition is provided, comprising:
[0095] Acquisition and processing module: used to obtain historical power data of wind farms and perform data cleaning;
[0096] Decomposition module: used to decompose historical wind power data into k components;
[0097] Component division module: used to perform component complexity analysis on the decomposed wind power subsequence components and divide them into high-frequency components and low-frequency components;
[0098] Model training and optimization module: used to train and optimize the established CLformer model and train the AESELM model;
[0099] Power prediction module: It is used to send the test sets of high-frequency and low-frequency components into the corresponding trained and optimized models for prediction, and to perform weighted summation on the measurement results based on the entropy weight method to obtain the final wind power combination prediction result.
[0100] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) The present invention proposes to use SVMD to decompose complex wind power data into a series of modal functions, each modal function represents a different frequency component in the wind power data; until all frequency components are extracted; this successive decomposition method can better capture the different frequency components in the wind power data, and facilitate and improve the accuracy of wind power prediction; (2) The present invention proposes to use FE to perform component complexity analysis on the wind power subsequence components, use the fuzzy entropy method to calculate each component, and divide each component into high-frequency and low-frequency signals according to the principle of the maximum difference in fuzzy entropy between adjacent components, so it is more in line with practical applications; (3) The present invention proposes that the CLformer prediction model enhances the model's ability to capture the local time dynamics of the sequence by introducing the local group autocorrelation (LGAC) mechanism, and uses the superposition of multiple LGAC layers to enable the model to capture multi-scale dependencies, which is particularly important for the prediction of high-frequency components, because high-frequency components often involve rapid changes and local patterns in the time series, thereby improving the model's prediction performance for high-frequency components; CLformer is superior to models using global autocorrelation mechanisms and self-attention in terms of efficiency and accuracy; (4) The present invention A method for predicting low-frequency components using AESELM is proposed. AESELM has a fast learning speed and is not easy to fall into local optimality. For a single hidden layer neural network, the input weights and biases can be randomly initialized and the corresponding output weights can be obtained, which improves the prediction accuracy and convergence speed of the model. This ability is very important for low-frequency components containing time series trends and periodic characteristics, and can effectively improve the prediction accuracy of low-frequency components. (5) The present invention proposes an IQIO algorithm that combines multiple improved strategies and adopts the Cubic chaotic mapping strategy population initialization method to obtain better optimization results. In the original QIOQ algorithm, The Gaussian random walk strategy is introduced in the position update. Through the guidance of the optimal individual, the population generates new individuals near the optimal individual, which is conducive to the rapid convergence of the algorithm. IQIO is used to optimize the number of potential factors and the number of hidden layers of the deep neural network in the matrix decomposition of the CLformer model, which can prevent overfitting and improve the generalization ability of the model. (6) The present invention proposes a weighting method based on the entropy weight method. Different from the traditional subjective weighting method, the entropy weight method does not rely on the subjective judgment of experts, but determines the weight by calculating the information entropy of the indicator of the data itself, making the determination of the weight more objective and effectively improving the accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0101] Figure 1 It is a schematic diagram of the method flow of the present invention.
[0102] Figure 2 It is a schematic diagram of the CLformer model structure of the present invention.
[0103] Figure 3 Schematic diagram of the IQIO algorithm of the present invention. DETAILED DESCRIPTION
[0104] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0105] like Figure 1 As shown, a wind power combination prediction method based on successive variational mode decomposition includes the following steps:
[0106] Step 1: Obtain historical power data of wind farms and perform data cleaning; cleaning includes processing, cleaning and modifying the data to eliminate errors, missing values, duplicate values, outliers, etc.
[0107] Step 2: Decompose the wind power data obtained in step 1 by SVMD. The steps are as follows:
[0108] Step 2.1, SVMD can decompose complex data into multiple components without setting the number of modes, and because SVMD has strong anti-noise ability, it is particularly suitable for processing nonlinear and non-stationary signals. The original wind power data is decomposed into two parts: k components and residual signals. The decomposition formula is as follows:
[0109] f(t)=u k (t)+f r (t) (1)
[0110] Among them, the residual signal f r (t) includes the unprocessed part f u (t) and the sum of the previously obtained components
[0111] Step 2.2, define the final decomposition formula:
[0112]
[0113] Where k represents the total number of components of the wind power data subsequence, u k (t) is the kth wind power data component, is the sum of the components of the wind power data subsequence obtained previously, f u (t) is the wind power data that has not been processed.
[0114] Step 3: Use the fuzzy entropy algorithm FE to perform a complexity analysis on the wind power subsequence components decomposed in step 2 and divide them into high-frequency and low-frequency components, and divide them into a training set and a test set;
[0115] The specific steps of separating into high-frequency and low-frequency components include:
[0116] Step 3.1: For the time series X (t) , the fuzzy entropy FE calculation method is as follows:
[0117]
[0118]
[0119]
[0120] Where m is the embedding dimension, is the similarity limit; X i , X j For the time series X (t) subsequence in; For X i , X j The values between them are the maximum values of their respective endpoint differences; r is the similarity; n is the gradient; is the constructor; S is X (t) length;
[0121] Step 3.2: For the components f1…f of SVMD decomposition k , where k is the total number of components, calculate its fuzzy entropy value, then compare the fuzzy entropy differences between adjacent components, and select the two components with the largest fuzzy entropy difference f i and f i+1 , the components f1~f i Defined as high-frequency component, the component f i+1 ~f k Defined as the low frequency component.
[0122] Step 4: Establish the CLformer model, such as Figure 2 As shown in the figure, the CLformer model is a lightweight model based on the Transformer architecture. The LGAC mechanism is proposed in the CLformer model, which calculates autocorrelation in equal-length sequence segments and enhances the model's ability to extract local time dynamics. By setting different sequence grouping lengths, LGAC can communicate between non-overlapping sequence segments and capture multi-scale time patterns. The high-frequency component training set obtained in step 3 is sent to the CLformer to train the model; the specific steps are:
[0123] Step 4.1: Convert the high-frequency time series of wind power X en ∈R L×m Input into the encoder and then transformed into X through the embedding layer en ∈R L×D ; Among them, X en It represents the matrix of high-frequency time series; R L×m represents a set of high-frequency real numbers with L rows and m columns, RL×D It represents a high-frequency real number set encoder with L rows and D columns;
[0124] Step 4.2: Encoder input is X en ∈R L×m , the encoder output is represents a real data set associated with high frequencies at N time steps, and D is the dimension of the hidden state; the encoder calculation process is as follows:
[0125] X t =AvgPool(Padding(X))
[0126] X s =XX t
[0127] X t =SeriesDecomp(X)
[0128]
[0129] in They represent the seasonal components obtained from the i-th time series decomposition component of the l-th layer, respectively, where l∈{1,…,}, TCB is the time convolution function, FeedForward is the feedforward neural network function, SeriesDecomp is the sequence decomposition function, AvgPool is the average function, Padding is the padding function, and X t is the result of decomposing the filled X sequence, X s For X and X t The difference between
[0130] Step 4.3: and In the input decoder, O is the prediction range; It represents the dataset of the initial time step when decoding the seasonal component in wind power forecasting; It represents the data set of the initial time step when decoding the trend period component in wind power forecasting; and Medium length The encoder input The seasonal and trend components are decomposed from the second half of ; the decoder calculation process is as follows:
[0131]
[0132] in i∈{1,2,3} represents the seasonal and trend components of the i-th time series decomposition component of the l-th layer; LGAC is the local grouped autocorrelation function; where l∈{1,...,M}, M is the number of decoder layers with the same decoder, ω l ∈R D ×n Represents the forecast trend component The sum of s Indicates that The dimension is converted from D to the projection matrix of the target dimension n; Sum and In projection Afterwards, the results are output as a time series of CLformer predictions.
[0133] Step 5: Establish the AESELM model. The AESELM model inherits the fast learning characteristics of ELM, but introduces an adaptive mechanism based on ELM, which means that the model can dynamically adjust its parameters according to the characteristics of the data, which can effectively improve the search ability and generalization ability of the model and effectively improve the prediction accuracy. Send the low-frequency component training set in step 3 to the AESELM to train the model; the steps are as follows:
[0134] Step 5.1: Map the low-frequency wind power data from the input space to the l-dimensional hidden layer feature space (AESELM feature space). The output of AESELM is:
[0135]
[0136] Among them, L is the number of hidden layer nodes, β i is the weight of the hidden layer and the output layer, h i (x) is the activation function; H is the hidden layer output matrix, β is the output weight matrix;
[0137] Step 5.2: Define the activation function as g(x) through the recorded training set, select the number of nodes L in the hidden layer, randomly generate input weights ωi and hidden layer bias i, calculate the hidden layer output matrix H, and obtain the wind power prediction result χ:
[0138]
[0139] Among them, H + is the generalized inverse matrix of H, N is the number of input wind power sequences, and T is the desired target matrix.
[0140] Step 6: Use Cubic chaotic map initialization and Gaussian random walk strategy to improve and optimize the quadratic interpolation optimization algorithm QIO to obtain the IQIO algorithm; use the quadratic interpolation optimization algorithm IQIO to optimize the hyperparameters of the CLformer model; Figure 3As shown, the specific steps are as follows:
[0141] Step 6.1, set the population size and number of iterations of the IQIO algorithm as well as the upper and lower limits of the search space;
[0142] Step 6.2: Improve the population initialization of the original QIO algorithm and use the good point set strategy to initialize the population of the QIO algorithm;
[0143]
[0144] p n (y) = {(L1y1,L2y2,...L n y n )},y=1,2,3...n
[0145]
[0146] Among them, a y is the lower dimension of the current dimension, b j is the upper dimension of the current dimension, γ i is the i-th individual, L j For the best point, p n (y) is the set of good points, Updated population.
[0147] Step 6.3, define the objective function as the mean absolute error (MAE) between the predicted value and the actual value of the wind power forecast event, and calculate the fitness value of the population through the objective function;
[0148] Step 6.4, individual update; search for the optimal hyperparameters of the CLformer model through different update methods. The hyperparameters include the number of potential factors in matrix decomposition and the number of hidden layers of the deep neural network, so that the average value of the absolute error between the predicted value and the actual value of the wind power forecast event is reduced;
[0149] First: Individual positions are updated using exploration strategies:
[0150]
[0151] ξ1=γ i+1 (m),γ rand1 (m),γ rand2 (m)
[0152]
[0153] w1=3n1b
[0154]
[0155]
[0156]
[0157] Where * represents the GIQ function of the number, γ rand represents individuals at different positions randomly selected from the group, ξ is the constructor, v i (m+1) indicates that at the m+1th iteration, individual γ i , w represents the exploration weight, n follows the standard normal distribution, b is the adaptive parameter, m is the number of iterations, M is the maximum number of iterations, e is the control parameter of b, p is a random number (0, 1), Represents the round function;
[0158] The valid value of GIQ function is: f(x) = x 2 -10cos(2πx)+10x∈[0.6,2].
[0159] Second: Individual location uses development strategy update method:
[0160]
[0161] ξ2=γ best (m),γ rand1 (m),γ rand2 (m)
[0162]
[0163]
[0164]
[0165] Introduce Gaussian random walk to replace the individual position update of the original QIO algorithm:
[0166]
[0167] σ=γ(m)-γ rand (m) (13)
[0168] Among them, pD represents a random integer in [1, d], Ub pD ,Lb pD are the upper and lower limits of the pDth dimension, ε i+1 has a mean of 0 and a variance of σ 2 A normally distributed random variable.
[0169] Step 6.5: Continuously update the current solution through algorithm iteration, and simultaneously update the global optimal solution;
[0170] Step 6.6: Determine whether the algorithm termination condition is reached through the given maximum number of iterations, and finally output the optimal hyperparameters of the CLformer model within the maximum number of iterations.
[0171] Step 7: Use the CLformer model and AESELM model trained and optimized in step (6) to predict the wind power of the wind farm, and perform weighted summation on the measurement results based on the entropy weight method to obtain the final prediction result.
[0172] Step 7.1: Send the high-frequency component test set to the trained and optimized CLformer model for prediction, and send the low-frequency component test set to the trained AESELM model for prediction to obtain the prediction results of each component; at the same time, the component results are weighted and summed based on the entropy weight method. The entropy weight method first standardizes the data:
[0173]
[0174] Among them l ij is the normalized value of the i-th component on the j-th index, x ij is the value of the i-th component at the j-th index, x max ( j ) is the maximum value on the jth index, x mi n( j ) is the minimum value on the jth index;
[0175] After standardizing each component, calculate the weight of each component under each indicator:
[0176]
[0177] where p ij It represents the weight of the i-th component under the j-th index;
[0178] Then calculate the information entropy and entropy weight of each indicator:
[0179]
[0180]
[0181] where e j 、w j are the information entropy and weight of the jth indicator, n and m are the number of components and the number of indicators, respectively. is a constant used to normalize entropy values;
[0182] Finally, the calculated entropy value is used to calculate the comprehensive score of each component:
[0183]
[0184] The final weight of each component is obtained by calculating the comprehensive score of each component, and the final wind power prediction result is obtained according to the weight.
[0185] The present invention provides a wind power combination prediction system based on successive variational modal decomposition, comprising:
[0186] Acquisition and processing module: used to obtain historical power data of wind farms and perform data cleaning;
[0187] Decomposition module: used to decompose historical wind power data into k components;
[0188] Component division module: used to perform component complexity analysis on the decomposed wind power subsequence components and divide them into high-frequency components and low-frequency components;
[0189] Model training and optimization module: used to train and optimize the established CLformer model and train the AESELM model;
[0190] Power prediction module: It is used to send the test sets of high-frequency and low-frequency components into the corresponding trained and optimized models for prediction, and to perform weighted summation on the measurement results based on the entropy weight method to obtain the final wind power combination prediction result.
Claims
1. A wind power combination prediction method based on successive variational mode decomposition, characterized in that: The steps include: Step 1: Obtain historical power data of the wind farm and perform data cleaning; Step 2: Decompose the historical wind power data into k components using SVMD; Step 3: Use the fuzzy entropy algorithm FE to perform a complexity analysis on the wind power subsequence components decomposed by SVMD and divide them into high-frequency components and low-frequency components, and divide them into a training set and a test set; Step 4: Establish a CLformer model, and send the high-frequency component training set obtained in step (3) into the CLformer model for model training; Step 5: Establish the AESELM model, and send the low-frequency component training set in step (3) into the AESELM for model training; Step 6: The quadratic interpolation optimization algorithm QIO is improved and optimized by using Cubic chaotic map initialization and Gaussian random walk strategy to obtain the IQIO algorithm; the hyperparameters of the CLformer model are optimized by using the quadratic interpolation optimization algorithm IQIO; Step 7: Send the high-frequency component test set data of step (3) into the trained and optimized CLformer model, and send the low-frequency component test set into the AESELM model to predict the wind power of the wind farm respectively, and perform weighted summation on the measurement results based on the entropy weight method to obtain the final prediction result.
2. The wind power combination prediction method based on successive variational modal decomposition according to claim 1 is characterized in that: The step 2 decomposes the input signal by SVMD, and the steps are as follows: Step 2.1: Decompose the original wind power data into two parts: k components and residual signal. The decomposition formula is as follows: f(t)=u k (t)+f r (t) (1) Among them, the residual signal f r (t) includes the unprocessed part f u (t) and the sum of the previously obtained components Step 2.2, define the final decomposition formula: Where k represents the total number of components of the wind power data subsequence, u k (t) is the kth wind power data component, is the sum of the components of the wind power data subsequence obtained previously, f u (t) is the wind power data that has not been processed.
3. The wind power combination prediction method based on successive variational modal decomposition according to claim 1 is characterized in that: The implementation process of step 3 is as follows: Step 3.1: For the time series X (t) , the fuzzy entropy FE calculation method is as follows: Where m is the embedding dimension, is the similarity limit; X i , X j For the time series X (t) subsequence in; For X i , X j The values between them are the maximum values of their respective endpoint differences; r is the similarity; n is the gradient; is the constructor; S is X (t) length; Step 3.2: For the components f1…f of SVMD decomposition k , where k is the total number of components, calculate its fuzzy entropy value, then compare the fuzzy entropy differences between adjacent components, and select the two components with the largest fuzzy entropy difference f i and f i+1 , the components f1~f i Defined as high-frequency component, the component f i+1 ~f k Defined as the low frequency component.
4. The wind power combination prediction method based on successive variational modal decomposition according to claim 1 is characterized in that: The step 4 inputs the high-frequency wind power subsequence output from the fuzzy entropy FE into the CLformer for prediction, and the implementation process is as follows: Step 4.1: Convert the high-frequency time series of wind power X en ∈R L×m Input into the encoder and then transformed into X through the embedding layer en ∈R L×D ; Among them, X en It represents the matrix of high-frequency time series; R L×m represents a set of high-frequency real numbers with L rows and m columns, R L×D It represents a high-frequency real number set encoder with L rows and D columns; Step 4.2: Encoder input is X en ∈R L×m , the encoder output is represents a real data set associated with high frequencies at N time steps, and D is the dimension of the hidden state; the encoder calculation process is as follows: in They represent the seasonal components obtained from the i-th time series decomposition component of the l-th layer, respectively, where l∈{1,…,}, TCB is the time convolution function, FeedForward is the feedforward neural network function, SeriesDecomp is the sequence decomposition function, AvgPool is the average function, Padding is the padding function, and X t is the result of decomposing the filled X sequence, X s For X and X t The difference between Step 4.3: and In the input decoder, O is the prediction range; It represents the dataset of the initial time step when decoding the seasonal component in wind power forecasting; It represents the data set of the initial time step when decoding the trend period component in wind power forecasting; and Medium length The encoder input The seasonal and trend components are decomposed from the second half of ; the decoder calculation process is as follows: in i∈{1,2,3} represents the seasonal and trend components of the i-th time series decomposition component of the l-th layer; LGAC is the local grouped autocorrelation function; where l∈{1,...,M}, M is the number of decoder layers with the same decoder, ω l ∈R D×n Represents the forecast trend component The sum of s Indicates that The dimension is converted from D to the projection matrix of the target dimension n; Sum and In projection Afterwards, the results are output as a time series of CLformer predictions.
5. The wind power combination prediction method based on successive variational modal decomposition according to claim 1 is characterized in that: The step 5 inputs the low-frequency wind power subsequence output from the fuzzy entropy FE into the AESELM model for prediction. The prediction process is as follows: Step 5.1: Map the low-frequency wind power data from the input space to the AESELM feature space. The output of AESELM is: Among them, L is the number of hidden layer nodes, β i is the weight of the hidden layer and the output layer, h i (x) is the activation function; H is the hidden layer output matrix, β is the output weight matrix; Step 5.2: Define the activation function as g(x) through the recorded training set, select the number of nodes L in the hidden layer, randomly generate input weights ωi and hidden layer bias i, calculate the hidden layer output matrix H, and obtain the wind power prediction result χ: Among them, H + is the generalized inverse matrix of H, N is the number of input wind power sequences, and T is the desired target matrix.
6. The wind power combination prediction method based on successive variational modal decomposition according to claim 1 is characterized in that: The step 6 is implemented as follows: Step 6.1, set the population size and number of iterations of the IQIO algorithm as well as the upper and lower limits of the search space; Step 6.2: Improve the population initialization of the original QIO algorithm and use the good point set strategy to initialize the population of the QIO algorithm; Among them, a y is the lower dimension of the current dimension, b j is the upper dimension of the current dimension, γ i is the i-th individual, L j For the best point, p n (y) is the set of good points, Updated population. Step 6.3, define the objective function as the mean absolute error (MAE) between the predicted value and the actual value of the wind power forecast event, and calculate the fitness value of the population through the objective function; Step 6.4, individual update: search for the optimal hyperparameters of the CLformer model through different update methods. The hyperparameters include the number of potential factors in matrix decomposition and the number of hidden layers of the deep neural network, so that the average value of the absolute error between the predicted value and the actual value of the wind power forecast event is reduced; Step 6.5: Continuously update the current solution through algorithm iteration, and simultaneously update the global optimal solution; Step 6.6: Determine whether the algorithm termination condition is reached through the given maximum number of iterations, and finally output the optimal hyperparameters of the CLformer model within the maximum number of iterations.
7. The method for predicting wind power events based on data enhancement according to claim 6, characterized in that: In step 6.4, the individual positions are updated using the exploration strategy: Where * represents the GIQ function of the number, γ rand represents individuals at different positions randomly selected from the group, ξ is the constructor, v i (m+1) indicates that at the m+1th iteration, individual γ i , w represents the exploration weight, n follows the standard normal distribution, b is the adaptive parameter, m is the number of iterations, M is the maximum number of iterations, e is the control parameter of b, p is a random number (0, 1), Represents the round function; The valid value of GIQ function is: f(x) = x 2 -10cos(2πx)+10x∈[0.6,2].
8. The method for predicting wind power events based on data enhancement according to claim 6, characterized in that: In step 6.4, the individual positions are updated using the development strategy: Introduce Gaussian random walk to replace the individual position update of the original QIO algorithm: Among them, pD represents a random integer in [1, d], Ub pD ,Lb pD are the upper and lower limits of the pDth dimension, ε i+1 has a mean of 0 and a variance of σ 2 A normally distributed random variable.
9. The wind power combination prediction method based on successive variational modal decomposition according to claim 1 is characterized in that: The implementation process of step 7 is as follows: The high-frequency component test set is sent to the trained and optimized CLformer model for prediction, and the low-frequency component test set is sent to the trained AESELM model for prediction to obtain the prediction results of each component; at the same time, the component results are weighted and summed based on the entropy weight method. The entropy weight method first standardizes the data: Among them l ij is the normalized value of the i-th component on the j-th index, x ij is the value of the i-th component at the j-th index, x max(j) is the maximum value on the jth index, x min(j) is the minimum value of the jth index; After standardizing each component, calculate the weight of each component under each indicator: where p ij It represents the weight of the i-th component under the j-th index; Then calculate the information entropy and entropy weight of each indicator: where e j 、w j are the information entropy and weight of the jth indicator, n and m are the number of components and the number of indicators, respectively. is a constant used to normalize entropy values; Finally, the calculated entropy value is used to calculate the comprehensive score of each component: The final weight of each component is obtained by calculating the comprehensive score of each component, and the final wind power prediction result is obtained according to the weight.
10. A wind power combination prediction system based on successive variational mode decomposition, characterized in that: include: Acquisition and processing module: used to obtain historical power data of wind farms and perform data cleaning; Decomposition module: used to decompose historical wind power data into k components; Component division module: used to perform component complexity analysis on the decomposed wind power subsequence components and divide them into high-frequency components and low-frequency components; Model training and optimization module: used to train and optimize the established CLformer model and train the AESELM model; Power prediction module: It is used to send the test sets of high-frequency and low-frequency components into the corresponding trained and optimized models for prediction, and to perform weighted summation on the measurement results based on the entropy weight method to obtain the final wind power combination prediction result.