Short-term wind power prediction method based on fine-grained feature extraction and model fusion

Abstract

The invention discloses a short-term wind power prediction method based on fine-grained feature extraction and model fusion. The method comprises the steps that original wind speed and wind power historical data are acquired and preprocessed, and a training set and a test set are divided; performing primary decomposition on the sequence by using SSA-VMD to obtain a plurality of IMF components and a residual sequence, and performing secondary decomposition on the residual sequence by EWT to obtain a residual subsequence; calculating an average influence value MIV between each component and the target power, screening high-correlation component fine-grained features, and constructing a multi-dimensional feature sequence; and an IA-KANTransform model is established, and training and prediction are carried out on the IA-KANTransform model. According to the method, the problems of non-stationarity and insufficient multi-scale feature extraction in wind power prediction are effectively solved, and the prediction precision and the generalization ability of the model are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power prediction, in particular to a short-term wind power prediction method based on fine-grained feature extraction and model fusion. BACKGROUND

[0002] With the aggravation of traditional energy shortage, developing renewable energy has become a global consensus. Wind power prediction is of great significance for achieving sustainable development, ensuring the stability of power systems and scheduling planning. Wind power series has significant randomness, nonlinearity, intermittency and non-stationarity, making accurate prediction a challenging task. Traditional wind power prediction methods mainly include physical methods and statistical methods. Physical methods usually rely on numerical weather prediction and basic parameters of wind farms, although they can reflect the variation of wind energy, but the cost of information acquisition is high and the calculation is complex. Statistical methods are based on historical observation data, such as autoregressive models, support vector machines and traditional neural networks, which can capture linear patterns, but have limited performance in handling non-linear and non-stationary data, making it difficult to balance local fluctuations and global trends.

[0003] In recent years, deep learning methods have shown strong performance in wind power prediction, especially the Transformer model, which is widely used due to its strong global feature capturing ability. However, the standard Transformer model still has shortcomings in wind power prediction, its feedforward neural network (FFN) usually uses a multilayer perceptron (MLP) with a fixed activation function, which is difficult to fully fit complex nonlinear relationships and local mutation features, affecting the prediction accuracy of non-stationary time series data. At the same time, single time or frequency domain decomposition methods are difficult to fully extract multi-scale information, and the residual signal still contains effective features that have not been analyzed.

[0004] Although existing wind power prediction methods have certain advantages in global feature extraction, they still have insufficient prediction performance when dealing with non-stationary and strongly random fluctuation wind power data. Specifically, these methods are difficult to balance overall trends and local fluctuations, prone to mode mixing, and the ability to extract local fine-grained features needs to be improved. Although the Transformer model has the ability to extract global features, its FFN relies on MLP with a fixed activation function, which is insufficient to fit complex nonlinearities and local mutations, affecting its prediction accuracy for non-stationary data.

[0005] Therefore, how to realize fine decomposition in wind power sequence modeling, fully extract multi-scale features, improve nonlinear modeling capability while considering local fluctuations and overall trends, ensure the stability of the model training process, and further improve the accuracy and precision of wind power prediction, and relieve gradient shock and improve stability, are the problems that need to be solved in current wind power prediction. Based on this, the present application proposes a short-term wind power prediction method based on fine-grained feature extraction and model fusion. The method has good stability and accuracy, and provides more solid and reliable support for realizing high-precision wind power prediction, sustainable development of energy systems, and stable operation of power systems. SUMMARY

[0006] To solve the problems existing in the prior art, the present application aims to provide a short-term wind power prediction method based on fine-grained feature extraction and model fusion, which can improve nonlinear modeling capability while considering local fluctuations and overall trends, and ensure the stability of the model training process.

[0007] To achieve the above-mentioned application purposes, the technical solutions adopted by the present application are as follows:

[0008] (1) Obtain original wind speed and wind power historical data;

[0009] (2) Perform abnormal value processing, missing value interpolation, and normalization processing on the historical data, and divide the training set and the test set;

[0010] (3) Use the SSA-VMD algorithm to decompose the wind speed and wind power sequence once to obtain a plurality of intrinsic mode functions (IMF) and a residual sequence, and use the EWT algorithm to decompose the residual sequence twice to obtain a residual sub-sequence;

[0011] (4) Calculate the average influence value (MIV) between each component and the target wind power, retain the high-correlation feature components, extract fine-grained features, and construct a multi-dimensional feature sequence;

[0012] (5) Construct an IA-KANTransformer prediction model, and use the training data set to train the wind power prediction model;

[0013] (6) Input the test set into the trained IA-KANTransformer wind power prediction model, and output the wind power prediction result.

[0014] Further, in step (1), long-time continuous wind speed and wind power historical data are obtained from the monitoring system of an actual wind farm, and the historical data are summarized and arranged to construct a historical data set for short-term wind power prediction.

[0015] Further, in the step (2), the historical data set in step (1) is preprocessed, including missing value filling, outlier processing, normalization and training set and test set division. The outliers in the historical data set are deleted, the missing data positions are replaced with the estimated values of the adjacent complete samples by using the nearest neighbor interpolation algorithm based on complete samples, and a complete historical data set is obtained; the processed historical data set is normalized by using the reversible instance normalization ReVIN method, and normalization and inverse normalization are independently performed on each input sequence instance to eliminate the distribution difference of the input sequence; the normalized data samples are divided into training set and test set, wherein 80% of the samples are used as training set and 20% of the samples are used as test set for subsequent model training and performance verification.

[0016] Further, in the step (3), the Sparrow Search Algorithm (SSA) is applied to adaptively optimize two key parameters of the Variational mode decomposition (VMD) algorithm, the mode number H and the penalty parameter α, to minimize the envelope entropy of the signal decomposition as the fitness function, and the sequence is decomposed by the VMD algorithm based on the optimal parameters H * 、α * , the decomposition result contains H IMF components and a residual sequence r(t), and the VMD minimizes the objective function under the constraint condition

[0017]

[0018] Wherein, f(t) represents the input signal, {u h} represents the hth mode component, {φ h} represents the hth center frequency, δ(t) represents the unit impulse function, and j represents the imaginary unit; the Lagrange multiplier λ and the penalty factor α are introduced to convert the constrained variation problem into an unconstrained variation problem, and the augmented Lagrange expression is obtained as:

[0019] L({u h},{w h},λ)

[0020]

[0021] The mode component and the center frequency set are continuously updated by using the alternating direction multiplier algorithm, and the components of different center frequencies are finally obtained;

[0022] ​Further, after obtaining the residual sequence, the empirical wavelet transform (EWT) is used to perform secondary decomposition on the residual sequence, the spectrum of the residual sequence is multiplied by each frequency band filter, and inverse Fourier transform is performed on the result to obtain the corresponding subsequence as:

[0023]

[0024] The residual sequence can be expressed as the superposition of all subsequences:

[0025]

[0026] Wherein, Indicates inverse Fourier transform, Indicates the spectrum of the residual sequence, Indicates the corresponding filter set.

[0027] Further, the method for optimizing the number of VMD modes H and the penalty parameter a based on SSA includes the following steps:

[0028] ① Initialize the sparrow population: set the population size and the maximum number of iterations, and randomly generate the position of each sparrow, which represents the number of VMD decomposition modes and the penalty factor;

[0029] ② Calculate the initial fitness: perform VMD decomposition according to the position of the sparrow, and use the minimum envelope entropy as the fitness evaluation index;

[0030] ③ Update the sparrow position: the finder updates the position based on the current optimal solution, the follower updates the position according to the difference, and the alert updates the position towards the safe zone;

[0031] ④ Update the fitness value and the best position: evaluate the fitness of the new position, and select the optimal sparrow as the new finder;

[0032] ⑤ When the number of iterations reaches the maximum, the algorithm terminates and the optimal parameters are adopted; otherwise, return to step ②.

[0033] Further, in step (4), the mean influence value MIV between each component and the target wind power is calculated by using the back propagation neural network BP, the MIV value is sorted according to the absolute value size, the high correlation component is retained, the fine-grained feature is extracted, and the multi-dimensional feature sequence is constructed.

[0034] Further, in the step (5), the single B-spline basis function is replaced by a mixed basis function of B-spline and Chebyshev polynomial to improve the KAN network, and an IA-KAN network is obtained; the IA-KAN network is used to replace the feedforward neural network FFN part of the Transformer model, and an IA-KAN Transformer wind power prediction model is constructed.

[0035] Further, the KAN network is based on the Kolmogorov-Arnold representation theorem, and a set of learnable one-dimensional basis functions is constructed to map the input variables to a high-dimensional feature space, and a one-dimensional function is used to approximate the complex multivariate nonlinear relationship through weighted combination, and the function form is represented as:

[0036]

[0037] wherein M represents the number of basis functions, w i represents a learnable coefficient, ψ i (·) represents the i-th one-dimensional basis function.

[0038] The mixed basis function based on B-spline and Chebyshev polynomial can be defined as:

[0039]

[0040] The high-dimensional representation of the IA-KAN network is:

[0041]

[0042] wherein represents the i-th B-spline basis function, T i (x) represents the i-th Chebyshev polynomial basis function, and β∈[0,1] represents a mixed weight coefficient for adaptively adjusting the contributions of B-spline and Chebyshev polynomial.

[0043] Further, the IA-KAN Transformer wind power prediction model has an overall network structure composed of an input embedding layer, a position encoding, a multi-head self-attention mechanism, an IA-KAN network, a residual connection, and layer normalization. The input multi-dimensional features are mapped to a high-dimensional space through linear embedding, and position encoding in the form of sine and cosine functions is added to retain sequence information. After the position encoding is added to the embedding vector, the input sequence with position perception characteristics is obtained as:

[0044] X pos =XW e +PE

[0045] wherein W edenotes an embedding matrix, PE denotes a position encoding matrix used to encode the position information of each element in the sequence, so that the model can capture the sequential relationship;

[0046] The multi-head attention mechanism captures the feature correlation of different subspaces in the sequence by calculating the query matrix Q, the key matrix K and the value matrix V of multiple attention heads, and the calculation formula of Q, K and V is as follows:

[0047] Q=X pos W Q , K=X pos W K , V=X pos W V

[0048] The global dependency relationship of the input sequence is modeled by calculating the similarity of Q and K and performing weighted summation, and the attention calculation formula is as follows:

[0049]

[0050] Wherein, d k represents the dimension of the key vector, W Q , W K , W V represent trainable weight matrices;

[0051] The multi-head attention output is obtained by splicing the linear mapping to obtain the final attention result:

[0052] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W 0

[0053] Wherein, h represents the number of heads, W 0 represents an output mapping matrix;

[0054] The residual connection and layer normalization are used to maintain the stability of network training and accelerate the convergence process, the IA-KAN network is introduced to perform adaptive nonlinear mapping on the input features, and the feature expression capability is enhanced, and the prediction result is output after being processed by the full connection layer and the activation function.

[0055] Further, in the step (6), the multi-dimensional feature sequence obtained in the step (4) is used to train the IA-KANTransformer wind power prediction model constructed in the step (5), the test set is input into the trained IA-KANTransformer model for prediction, and the prediction value is processed by inverse normalization to obtain the prediction value under the actual power scale.

[0056] The beneficial effects of the present application are:

[0057] Compared with the existing wind power prediction method, the present application has significant advantages and improvements, which are embodied in the following aspects:

[0058] (1) Multi-layer fine-grained feature extraction

[0059] Compared with the existing prediction method relying only on single signal decomposition method for feature extraction, the present application adopts a multi-level fine-grained decomposition strategy based on SSA-VMD and EWT, taking into account the overall trend and local fluctuations, layer by layer mining and fusing the subtle local features in the historical data, fully separating different frequencies in the historical data, relieving the modal aliasing problem, realizing multi-scale expression of features, and improving the recognition and understanding ability of the model to small differences and complex structures.

[0060] (2) Hybrid basis function improved KAN

[0061] Compared with the existing KAN network using a single B-spline basis function, the present application constructs a hybrid basis function based on B-spline and Chebyshev polynomial, replacing the B-spline basis function in the traditional KAN network, so that the network not only retains the good local approximation ability of B-spline, but also further utilizes the advantages of Chebyshev polynomial in global fitting and numerical stability, improves the expression ability of the model to complex nonlinear relationships, enhances the robustness in noisy environment, and effectively alleviates the gradient shock and overfitting problems caused by a single basis function.

[0062] (3) IA-KAN enhanced Transformer

[0063] Compared with the existing wind power prediction model using FFN structure, the present application combines IA-KAN network and Transformer model, uses IA-KAN network to replace the FFN part of Transformer model, constructs IA-KAN Transformer model, so that the network not only retains the advantages of the original Transformer in global feature extraction, but also significantly enhances the fitting ability of the model to complex nonlinear relationships and local mutations, realizes more accurate and stable modeling effect.

[0064] (4) Prediction accuracy and robustness improvement

[0065] Compared with the existing prediction method relying on single feature extraction or traditional neural network structure, the present application significantly improves the accuracy and stability of wind power prediction by multi-level fine-grained decomposition, multi-scale feature construction, and combining IA-KAN network to enhance the nonlinear modeling ability of Transformer model, while enhancing the robustness to non-stationary, nonlinear and noisy interference data, providing reliable theoretical support and technical support for practical application. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 For the overall flowchart of the present application.

[0067] Figure 2 For the SSA optimized VMD parameter flowchart of the present application.

[0068] Figure 3 For the IA-KAN network flowchart of the present application.

[0069] Figure 4 For the IA-KAN Transformer model flowchart of the present application. DETAILED DESCRIPTION

[0070] In order to illustrate the content of the present application, the following examples will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be pointed out that the examples are preferred schemes of the present application, and are intended to illustrate the implementation conditions that can be used to realize the present application, rather than to limit the experimental conditions.

[0071] The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings of the present application specification, and the examples are only used to explain the present application, and are not used to limit the scope of the present application.

[0072] The present application provides a short-term wind power prediction method based on fine-grained feature extraction and model fusion, the method of the present application adopts SSA-VMD decomposition method to decompose wind speed and wind power sequence once, EWT algorithm to decompose residual sequence twice, calculate the average influence value MIV between each component and target wind power, retain high correlation feature component, extract fine-grained feature, construct hybrid basis function based on B-spline and Chebyshev polynomial, improve KAN network, establish IA-KAN network, replace the FFN part of the Transformer model with IA-KAN network, and construct IA-KANTransformer wind power prediction model, so as to improve the prediction accuracy of the model.

[0073] The short-term wind power prediction method based on fine-grained feature extraction and model fusion of the present application includes the following steps, such as Figure 1 :

[0074] Step S1: Obtain original wind speed and wind power historical data.

[0075] The long-time continuous wind speed and wind power historical data are obtained from the monitoring system of the actual wind farm, the historical data contains the wind speed change information and the corresponding power output of the wind turbine under different meteorological conditions and operating conditions, by summarizing and arranging the data, the historical data set for short-term wind power prediction is constructed.

[0076] In this embodiment, the experimental subject is an actual dataset from a wind farm in East China.

[0077] Step S2: Perform outlier handling, missing value imputation, and normalization on the historical dataset, and divide it into training and test sets.

[0078] S2-1: Remove outliers from the historical dataset. Using a nearest neighbor imputation algorithm based on complete samples, replace the missing values ​​at the locations of the missing data with the estimated values ​​of their neighboring complete samples to obtain the complete historical dataset. Assume this historical dataset is X = {x1, x2, ..., x...} n},x i ∈R m For any sample x with missing values i Define its missing features as Its interpolation value is expressed as:

[0079]

[0080] in, N represents the missing values ​​after imputation. W (i) represents the sample x i The most recent W complete sample sets, This represents the feature value corresponding to adjacent complete samples.

[0081] S2-2: The processed historical dataset is normalized using the ReVIN method. Normalization and inverse normalization are performed independently for each input sequence instance to eliminate the influence of different units on the analysis results and prevent variables with larger units from dominating the analysis. The mean and standard deviation are calculated separately for each time series sample, transforming the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. The normalization formula is:

[0082]

[0083] Where X represents the original data points, μ represents the sequence mean, and σ represents the sequence standard deviation. norm This represents the normalized data points.

[0084] S2-3: Divide the data samples into training set and test set for subsequent modeling, with 80% of the samples used as training set and 20% of the samples used as test set.

[0085] Step S3: The normalized data is decomposed once using the SSA-VMD decomposition method, and the residual sequence is decomposed a second time using the EWT algorithm.

[0086] S3-1: VMD is an adaptive and non-recursive signal decomposition algorithm, which constructs a variational constraint problem by determining the number of modal decomposition, realizes the effective separation of IMF component and the frequency domain division of signal, and obtains the optimal solution of variational problem. In the VMD algorithm, it is assumed that each mode is a limited bandwidth with center frequency, and the variational problem is described as seeking H modal functions, so that the sum of the estimated bandwidth of each mode is minimized, and the constraint condition is that the sum of each mode is equal to the input signal. Assuming that the input signal is f(t), the constraint expression of using VMD for signal decomposition is defined as follows:

[0087]

[0088] Where, {u h} represents the hth modal component, {φ h} represents the hth center frequency, δ(t) represents the unit impulse function, and j represents the imaginary unit.

[0089] To solve the optimization problem, the Lagrange multiplier λ and the penalty factor α are introduced, the constrained variational problem is converted into an unconstrained variational problem, and the augmented Lagrange expression is obtained as:

[0090]

[0091] The alternating direction multiplier algorithm is used to continuously update the modal component and center frequency set, and finally the components with different center frequencies are obtained.

[0092] The SSA is an intelligent optimization algorithm that simulates the foraging behavior of sparrow population. The sparrow population can be divided into discoverers, followers and guards. The discoverer can actively search for food and guide the movement of the group. The follower needs to find food by imitating the position of the discoverer. The guard is responsible for monitoring the environment. When the sparrow perceives a threat, i.e. the warning value is greater than the safety value, it will move away from the dangerous area to find food elsewhere. Assuming that the number of sparrow population is n, the position update formula of the discoverer in each iteration process is:

[0093]

[0094] Where, represents the j-dimensional position of the ith sparrow in the t+1th iteration, iter max represents the maximum number of iterations, R2 represents the alarm value, ST represents the safety threshold, and P represents the standard normal distribution parameter. When R2≥ST, it means that the sparrow discovers the predator and sends an alarm, and the population moves to the safety zone. When R2<ST, it means that it is safe to search for food. When the discoverer finds better food, the follower leaves the current position to compete for food.

[0095] The position update formula of the follower is:

[0096] The position update formula of the follower is:

[0097] where X p represents the best position, X worst represents the worst position, A represents a random matrix, and C represents an all-one vector. When i > q / 2, it means that the follower at time i has no food and needs to forage elsewhere to obtain more energy.

[0098] The position update formula of the alarm is:

[0099]

[0100] where X best represents the global best position, fit g represents the global optimal fitness, fit w represents the global worst fitness, fit i represents the fitness of the ith sparrow, b represents the step control parameter, v represents the direction of sparrow movement, and ζ represents a parameter to prevent the denominator from being zero. When fit i > fit g , the sparrow is in a non-optimal position and is vulnerable to attack. When fit i = fit g , the sparrow discovers danger and will move to a place close to itself.

[0101] The method for optimizing the number of VMD modalities H and the penalty parameter α based on SSA includes the following steps, as Figure 2 :

[0102] ① Initialize the sparrow population. Set the population size and the maximum number of iterations, and randomly generate the position of each sparrow, which represents the number of modalities of VMD decomposition and the penalty factor.

[0103] ② Calculate the initial fitness. Perform VMD decomposition according to the position of the sparrow, and use the minimum envelope entropy as the fitness evaluation index.

[0104] ③ Update the sparrow position. The discoverer updates the position based on the current optimal solution, the follower updates the position according to the difference, and the alarm tends to the safe zone to update the position.

[0105] ④ Update the fitness value and the best position. Evaluate the fitness of the new position, and select the optimal sparrow as the new discoverer.

[0106] ⑤ When the number of iterations reaches the maximum, the algorithm terminates and adopts the optimal parameters; otherwise, return to ②.

[0107] S3-2: EWT is a data-driven time-frequency analysis method, aiming at multi-scale decomposition of signals, and its core idea is to adaptively construct a set of band-pass filters according to the spectral characteristics of the signal, which can effectively decompose the signal according to its spectral characteristics. Unlike traditional Fourier transform or wavelet transform, EWT does not rely on fixed basis functions, but dynamically constructs adaptive filters according to the actual spectral characteristics of the signal.

[0108] The residual sequence is decomposed twice by using the empirical wavelet transform EWT, the spectrum of the residual sequence is adaptively divided into N continuous frequency bands, a band-pass wavelet filter is constructed on each frequency band, the spectrum of the residual sequence is multiplied by each frequency band filter, and the inverse Fourier transform is performed on the result to obtain the corresponding subsequence as:

[0109]

[0110] The residual sequence can be represented as the superposition of all sub-sequences:

[0111]

[0112] wherein, denotes the inverse Fourier transform, denotes the spectrum of the residual sequence, denotes the corresponding filter bank.

[0113] Step S4: MIV is an index used to measure the influence of each feature on the target output, and its absolute value represents the influence of the feature on the model output, and the larger the value, the stronger the correlation. The average influence value MIV between each component and the target wind power is calculated by using the back propagation neural network BP, and the calculation formula is:

[0114]

[0115] wherein, N represents the number of samples, and F(·) represents the BP model output. According to the absolute value size, sorting and screening are performed, high correlation components are retained, fine-grained features are extracted, and a multi-dimensional feature sequence containing multi-scale information is formed.

[0116] Step S5: a hybrid basis function is constructed based on B-spline and Chebyshev polynomial, KAN network is improved, IA-KAN is established, IA-KAN network is used to replace the FFN part of the Transformer model, and IA-KANTransformer wind power prediction model is constructed.

[0117] S5-1: The KAN network in this embodiment is based on the Kolmogorov-Arnold representation theorem, which shows that any continuous multivariate function can be represented as a finite sum of univariate continuous functions. By constructing a set of learnable univariate basis functions, the input variables are mapped to a high-dimensional feature space, and the complex multivariate nonlinear relationship is approximated by the weighted combination of univariate functions, and its function form is expressed as:

[0118]

[0119] where M represents the number of basis functions, w i represents a learnable coefficient, ψ i (·) represents the i-th univariate basis function.

[0120] The univariate basis function is usually represented by the B-spline basis function, which is a piecewise polynomial function commonly used for data fitting and curve modeling, and its calculation is recursive. When the order p = 0, i.e. the zero-order B-spline basis function, it is defined as:

[0121]

[0122] The function takes the value 1 in the corresponding node interval and 0 in other intervals. When p > 0, the recursive calculation of the B-spline basis function is expressed as:

[0123]

[0124] where u i represents the i-th node position in the node vector.

[0125] Chebyshev polynomials are a kind of orthogonal polynomials commonly used in numerical analysis and function approximation, which have good numerical stability and global approximation characteristics. The first kind of Chebyshev polynomials can be expressed as:

[0126] T i (x) = cos(iarccos(x)), x ∈ [-1, 1]

[0127] satisfy the following recursive relationship:

[0128] T0(x) = 1, T1(x) = x, T i+1 (x) = 2xT i (x) - T i-1 (x)

[0129] The mixed basis function based on B-spline and Chebyshev polynomial can be defined as:

[0130]

[0131] For example, Figure 3As shown, the input features are mapped using the hybrid basis functions, and the high-dimensional representation of the IA-KAN network is:

[0132]

[0133] wherein, denotes the i-th B-spline basis function, T i (x) denotes the i-th Chebyshev polynomial basis function, and β ∈ [0, 1] denotes the hybrid weight coefficient, which is used to adaptively adjust the contributions of the B-spline and the Chebyshev polynomial.

[0134] S5-2: Transformer is a deep learning model proposed by Vaswani et al. in 2017, mainly used for processing sequence data. This model abandons the traditional structure of recurrent neural network (RNN) and convolutional neural network (CNN), and adopts a self-attention mechanism and parallel computing method. The Transformer model mainly consists of an input embedding layer, position encoding, multi-head self-attention mechanism, feedforward neural network, and residual connection and layer normalization. The input sequence X is first mapped to a high-dimensional feature space through the embedding layer, and position encoding in the form of sine and cosine functions is added to preserve sequence information. The position encoding formula is:

[0135]

[0136] wherein, pos represents the position of the sequence element, 2i represents the even dimension, 2i+1 represents the odd dimension, d model represents the dimension of the embedding vector. After adding the position encoding to the embedding vector, the input sequence with position awareness characteristics is obtained as:

[0137] X pos =XW e +PE

[0138] wherein, W e represents the embedding matrix, and PE represents the position encoding matrix, which is used to encode the position information of each element in the sequence, so that the model can capture the sequential relationship.

[0139] The multi-head attention mechanism captures the feature correlation of different subspaces in the sequence by calculating the query matrix Q, the key matrix K, and the value matrix V of multiple attention heads. The calculation formulas of Q, K, and V are:

[0140] Q=X pos W Q , K=X pos W K , V=X pos W V

[0141] The global dependence of the input sequence is modeled by calculating the similarity of Q and K and performing weighted summation, and the attention calculation formula is:

[0142]

[0143] Where d k represents the dimension of the key vector, W Q , W K , W V represent trainable weight matrices.

[0144] The multi-head attention output is obtained by concatenating the linear mapping to obtain the final attention result:

[0145] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W 0

[0146] Where h represents the number of heads, W 0 represents the output mapping matrix.

[0147] After multi-head attention, residual connection and normalization processing are performed. The standard FFN is used for nonlinear mapping and dimension transformation of the features at each time step, and the calculation formula is:

[0148] FFN(x)=σ(xW1+b1)W2+b2

[0149] Where σ represents the activation function, W1 and W2 represent the weight matrix, and b1 and b2 represent the bias vector.

[0150] S5-4: The IA-KAN Transformer prediction model is a model constructed by replacing the FFN part with the IA-KAN network based on the standard Transformer model. This model combines the strong global modeling capability of the Transformer and the adaptive nonlinear mapping capability of the IA-KAN, and can more accurately depict the multi-scale dynamic characteristics of the wind power sequence. As shown in Figure 4As shown, the overall network structure of the IA-KANTransformer is composed of input embedding layer, position encoding, multi-head self-attention mechanism, IA-KAN network, residual connection and layer normalization. First, the input multi-dimensional features are mapped to a high-dimensional space through linear embedding, and position encoding in the form of sine and cosine functions is added to retain sequence information; second, the multi-head attention mechanism is introduced, the features of different subspaces in the sequence are captured respectively by calculating multiple attention heads and attention weights, the global dependence is extracted, and residual connection and layer normalization are added to keep the training stable; then the IA-KAN network is used to realize the adaptive nonlinear mapping of input features; finally, the prediction result is output through the full connection layer and the activation function.

[0151] Step S6: The IA-KANTransformer wind power prediction model constructed in step S5 is trained using the multi-dimensional feature sequence obtained in step S4, and the model parameters are optimized. After training, the test set is input into the trained IA-KANTransformer model for prediction, and the predicted value is processed by inverse normalization to obtain the predicted value in the actual power scale. The root mean square error RMSE, the mean absolute error MAE and the mean absolute percentage error MAPE are used as the error evaluation indexes of the prediction model, and the specific formula is:

[0152]

[0153] Wherein, S represents the number of samples, y i represents the true value, represents the predicted value.

[0154] The above examples are only used to illustrate the technical solutions of the present application and not to limit it, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A short-term wind power prediction method based on fine-grained feature extraction and model fusion, characterized in that, Includes the following steps: (1) Obtain historical data on raw wind speed and wind power; (2) Perform outlier processing, missing value imputation, and normalization on historical data, and divide the data into training and test sets; (3) The wind speed and wind power sequences are decomposed once using the SSA-VMD algorithm to obtain multiple intrinsic mode functions (IMFs) and a residual sequence. The residual sequence is then decomposed twice using the EWT algorithm to obtain residual subsequences. (4) Calculate the average influence value (MIV) between each component and the target wind power, retain the highly correlated feature components, extract fine-grained features, and construct a multi-dimensional feature sequence. (5) Construct the IA-KANTransformer prediction model and train the wind power prediction model using the training dataset; (6) Input the test set into the trained IA-KANTransformer wind power prediction model and output the wind power prediction results.

2. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 1, characterized in that, In step (1), historical data of wind speed and wind power collected continuously over a long period of time are obtained from the monitoring system of the actual wind farm. By summarizing and organizing the historical data, a historical dataset for short-term wind power prediction is constructed.

3. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 1, characterized in that, In step (2), outliers in the historical dataset are deleted, and the nearest neighbor interpolation algorithm based on complete samples is used to replace the missing values ​​at the missing data locations with the estimated values ​​of the adjacent complete samples to obtain the complete historical dataset. The processed historical dataset is normalized using the ReVIN method. Each input sequence instance is independently normalized and denormalized to eliminate the distribution differences of the input sequences. The normalized data samples are divided into training and test sets for subsequent model training and performance verification.

4. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 1, characterized in that, In step (3), the Sparrow Optimization Algorithm (SSA) is applied to adaptively optimize the two key parameters of Variational Mode Decomposition (VMD): the number of modes H and the penalty parameter α. The fitness function is to minimize the envelope entropy after signal decomposition, based on the optimal parameter H. * α * The VMD algorithm decomposes the sequence once, and the decomposition result contains H IMF components and a residual sequence r(t). VMD operates under constraints... The objective function to be minimized is: Where f(t) represents the input signal, {u h } represents the h-th modal component, {φ h Let} denote the h-th center frequency, δ(t) denote the unit impulse function, and j denote the imaginary unit; by introducing the Lagrange multiplier λ and the penalty factor α, the constrained variational problem is transformed into an unconstrained variational problem, yielding the augmented Lagrange expression: The modal components and center frequency set are continuously updated using the alternating direction multiplier algorithm, and finally the components with different center frequencies are obtained. After obtaining the residual sequence, the empirical wavelet transform (EWT) is further used to perform a secondary decomposition of the residual sequence. By multiplying the spectrum of the residual sequence by the filters of each frequency band and performing an inverse Fourier transform on the result, the corresponding subsequences are obtained as follows: The residual sequence can be represented as a superposition of all subsequences: in, Indicates the inverse Fourier transform. Represents the spectrum of the residual sequence. This indicates the corresponding filter bank.

5. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 4, characterized in that, The method for optimizing the number of VMD modes H and the penalty parameter α based on SSA includes the following steps: ① Initialize the sparrow population: Set the population size and maximum number of iterations, and randomly generate the position of each sparrow, which represents the number of modes in the VMD decomposition and the penalty factor; ② Calculate the initial fitness: Perform VMD decomposition based on the sparrow's position, and use the minimum envelope entropy as the fitness evaluation index; ③ Update the sparrow's position: The discoverer updates the position based on the current optimal solution, the follower updates the position based on the difference, and the vigilant moves towards the safe zone to update the position; ④ Update fitness value and optimal location: Evaluate the fitness of the new location and select the best sparrow as the new discoverer; ⑤ When the number of iterations reaches the maximum, the algorithm terminates and adopts the optimal parameters; otherwise, return to step ②.

6. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 1, characterized in that, In step (4), the average influence value (MIV) between each component and the target wind power is calculated using a backpropagation neural network (BP). The components are sorted according to their absolute values, and feature selection is performed. Highly correlated components are retained, and fine-grained features are extracted to form a multidimensional feature sequence.

7. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 1, characterized in that, In step (5), the KAN network is improved by replacing the single B-spline basis function with a mixed basis function of B-spline and Chebyshev polynomial to obtain the IA-KAN network; the feedforward neural network FFN part of the Transformer model is replaced by the IA-KAN network to construct the IA-KANTransformer wind power prediction model.

8. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 7, characterized in that, The KAN network is based on the Kolmogorov-Arnold representation theorem. It maps input variables to a high-dimensional feature space by constructing a set of learnable univariate basis functions, and approximates complex multivariate nonlinear relationships through a weighted combination of univariate functions. Its functional form is expressed as follows: Where M represents the number of basis functions, w i ψ represents the learnable coefficient. i (·) represents the i-th unary basis function. The mixed basis function based on B-splines and Chebyshev polynomials can be defined as: By mapping the input features using hybrid basis functions, the high-dimensional representation of the IA-KAN network is as follows: in, Let T represent the basis function of the i-th B-spline. i (x) represents the i-th Chebyshev polynomial basis function, and β∈[0,1] represents the mixed weight coefficient, which is used to adaptively adjust the contributions of B-spline and Chebyshev polynomial.

9. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 7, characterized in that, The overall network structure of the IA-KANTransformer wind power prediction model consists of an input embedding layer, position encoding, a multi-head self-attention mechanism, an IA-KAN network, residual connections, and layer normalization. The multidimensional features of the input are mapped to a high-dimensional space through linear embedding, and position encoding in the form of sine and cosine functions is added to preserve sequence information. After adding the position encoding to the embedding vector, the input sequence with position-aware characteristics is obtained as follows: X pos =XW e +PE Among them, W e PE represents the position encoding matrix, which is used to encode the position information of each element in the sequence, enabling the model to capture the sequential relationship; The multi-head attention mechanism captures feature associations in different subspaces of a sequence by calculating the query matrix Q, key matrix K, and value matrix V of multiple attention heads. The formulas for calculating Q, K, and V are as follows: Q=X pos W Q ,K=X pos W K ,V=X pos W V By calculating the similarity between Q and K and performing a weighted sum, the global dependency relationship of the input sequence is modeled. The attention calculation formula is as follows: Where, d k W represents the dimension of the key vector. Q W K W V This represents a trainable weight matrix; The multi-head attention output is spliced ​​with a linear mapping to obtain the final attention result: MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W 0 Where h represents the number of heads, W 0 Indicates the output mapping matrix; The residual connections and layer normalization are used to maintain the stability of network training and accelerate the convergence process. The IA-KAN network is introduced to perform adaptive nonlinear mapping on the input features, which enhances the feature representation ability. After processing by fully connected layers and activation functions, the prediction results are output.

10. The short-term wind power prediction method based on fine-grained feature extraction and model fusion according to claim 1, characterized in that, In step (6), the multidimensional feature sequence obtained in step (4) is used to train the IA-KANTransformer wind power prediction model constructed in step (5). The test set is input into the trained IA-KANTransformer model for prediction. The output prediction value is processed by inverse normalization to obtain the prediction value under the actual power scale.

Images