Interpretable short-term wind power generation prediction method based on Kolmogorov-Arnold network

Abstract

The invention discloses an interpretable short-term wind power generation prediction method based on a Kolmogorov-Arnold network, and relates to the field of wind power generation power prediction and frequency domain modeling analysis of a power system, and the method comprises the steps: carrying out the multi-source data collection of target power generation equipment in a historical time range, constructing a prediction input feature set, and carrying out the calculation of a prediction input feature set; acquiring a generation power time sequence and an environment correlation characteristic thereof; and according to the constructed input and output samples, constructing an initial KAN network and carrying out full-amount training. The method can provide mathematical model support for modeling, prediction and characteristic analysis of the power generation process of the wind power station. By introducing a symbolized activation function structure and a network pruning mechanism, a function relationship between input characteristics and power generation output can be accurately identified under a limited sample condition, and a prediction model with a clear mathematical analysis form is extracted, so that quantitative modeling and interpretable analysis of new energy power prediction are realized, and the prediction efficiency is improved. And the expression accuracy and engineering applicability of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of wind power generation prediction and frequency domain modeling analysis in power systems, and in particular to an interpretable short-term wind power generation prediction method based on the Kolmogorov-Arnold network. Background Technology

[0002] With the rapid integration of new energy sources, such as wind power and solar power, into the new power system, power production is exhibiting high volatility and uncertainty, and the dynamic characteristics of system operation are undergoing profound changes. Since wind power is generally connected to the grid through power electronic devices, system stability, dispatch security, and energy management face greater challenges. Therefore, establishing a high-precision, fast-response, and well-interpretable short-term wind power generation forecasting technology is of great significance for improving the real-time dispatching capabilities and operational transparency of the power system.

[0003] Currently, short-term wind power forecasting methods mainly include physical modeling methods, statistical learning methods, and deep learning methods. Physical modeling methods rely on external input variables such as meteorology, and while they can reflect the physical mechanisms of wind power generation, they suffer from problems such as complex modeling and poor universality. Statistical learning methods, such as ARIMA and support vector regression, have certain predictive capabilities under small sample conditions, but struggle to capture nonlinear feature changes. Deep learning methods, through the construction of multi-layered network structures, possess strong fitting capabilities and have achieved good results in multiple prediction scenarios; however, their model structure exhibits "black box" characteristics, making it difficult to meet the power industry's urgent needs for interpretability and reliability.

[0004] Therefore, it is necessary to develop a power generation prediction method that combines high accuracy and good interpretability to meet the practical requirements of complex power system operation for model reliability and transparency. Summary of the Invention

[0005] The purpose of this invention is to solve the problems mentioned in the background section by proposing an interpretable short-term wind power generation prediction method based on the Kolmogorov-Arnold network.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] Interpretable short-term wind power generation prediction methods based on Kolmogorov-Arnold networks include:

[0008] S1: Collect multi-source data from the target power generation equipment within a historical time range, construct a predictive input feature set, and obtain the power generation time series and its environmental correlation characteristics;

[0009] S2: Based on the constructed input and output samples, build an initial KAN network and perform full training to obtain a basic model that fits the predicted relationship;

[0010] S3: Introduce L1 norm and entropy regularization constraints to prune and sparsify the trained KAN network, thereby realizing the sparse representation structure of the KAN network.

[0011] S4: Based on the sparse training results, structural pruning is performed on low-response nodes to construct a KAN network with structural readability.

[0012] In a preferred embodiment of the present invention, in step S1, multi-source data is collected from the target power generation equipment within a historical time range. Specifically, for any target power generation equipment, multi-source feature variables and power output data during its operation are collected to obtain raw data. The raw data is then preprocessed and uniformly encoded to construct a prediction input-output dataset for KAN modeling, including:

[0013] S11: Historical Operation Data Acquisition: For the target power generation equipment to be modeled, historical operation data is collected under steady-state conditions during grid-connected operation, including environmental characteristic parameters and time characteristic information; at the same time, the actual power generation value at the corresponding moment is recorded to form a structured historical data sample set;

[0014] S12: Data Preprocessing and Standardization: All feature variables are preprocessed uniformly using the Z-Score standardization method to transform the j-th feature value of the i-th sample into a standardized feature value. Outliers are detected based on the 3σ principle and corrected using linear interpolation. For time features, sine and cosine functions are used to encode the time variable t into sin(2πt / T) and cos(2πt / T) forms, where T is the period. After processing, a standardized input feature vector and corresponding output target are formed.

[0015] S13: Input-output sample pair construction: Based on the standardized time series data, sample pairs of input feature vectors and corresponding output target values ​​are constructed according to the time order. Specifically, a fixed-length historical observation sequence is set as the input feature, and several consecutive time moments are extracted to form the input vector. The power generation data corresponding to a specified time step after the end of the selected input sequence is used as the output target. In this way, the complete dataset is traversed to construct N input-output sample pairs in sequence, and finally a training dataset D is formed for subsequent modeling.

[0016] S14: Dataset Partitioning and Format Encapsulation: The constructed training dataset is divided into training, validation, and test sets in chronological order and converted into a data format acceptable to the KAN network. All input feature vectors must ensure dimensionality consistency and retain normalization parameters. Output power data is used to construct target prediction values, ultimately forming a complete input-output data structure.

[0017] In a preferred embodiment of the present invention, step S2 involves constructing an initial KAN network and performing full training based on the constructed input and output samples, including:

[0018] S21: Network structure initialization: Based on the Kolmogorov-Arnold theorem's multidimensional function decomposition theory, power generation prediction is modeled as a combinatorial optimization problem of multi-layer one-dimensional nonlinear functions;

[0019] S22: Network parameter training: Based on the training dataset D, construct the Kolmogorov-Arnold network prediction function and its parameter set θ to be optimized. Define the loss function with the goal of minimizing the mean square error, and use the finite memory quasi-Newton method to iteratively solve the optimization problem.

[0020] S23: Dynamic Raster Refinement: During training, the spline activation function within the KAN is refined by raster points according to a preset step interval. In the initial stage, the activation function is constructed based on a coarse grid G1. Subsequently, to meet the complex response requirements within the activation region, a refined high-resolution grid G2 is introduced to reconstruct a finer spline activation function, where G2 > G1.

[0021] S24: Training convergence and error monitoring: Monitor the changes in the training loss curve in real time to confirm that the model has reached a stable convergence state.

[0022] In a preferred embodiment of the present invention, step S3 involves pruning and sparsifying the trained KAN network, including:

[0023] S31: Activation Function Response Statistics and Norm Constraint Construction: Statistical modeling of the response amplitudes of neurons in each layer is performed, assuming... The layer activation output matrix is Where Np is the number of samples. Given the number of active nodes in this layer, the L1 norm is defined as:

[0024] ;

[0025] S32: Construction of Normalized Entropy Regularization Term and Modeling of Distribution Balance: To standardize the distribution of activation path responses among nodes, an entropy regularization mechanism is introduced, defining the normalized response of the j-th neuron node as:

[0026] ;

[0027] The normalized information entropy of this layer is calculated as follows:

[0028] ;

[0029] S33: Construction of the sparsified objective function and full training: Based on the response intensity norm and normalized entropy term in step S2, construct an objective function that includes a structure regularization term:

[0030] ;

[0031] in, For the original predicted loss, , The weight hyperparameter for the regularization term represents the degree to which the network adjusts for sparsity and distribution.

[0032] In a preferred embodiment of the present invention, step S4 involves structural pruning of low-response nodes based on the sparsity training results, including:

[0033] S41: Activation Path Pruning and Structure Compression Based on Sparse Responses: After completing sparse training with L1 norm and entropy constraints, automatic pruning is performed on neuronal pathways with weak response amplitudes and low entropy distribution concentration. Specifically, for the i-th node, if its average activation response satisfies: If θ1 is the response threshold, it is determined to be a low contribution path, corresponding to redundant feature expression. Such nodes are removed from the network connection graph, and the weight matrix of adjacent layers is updated synchronously.

[0034] S42: KAN Network Reconstruction and Readable Structure Generation Based on Pruning Results: For the KAN network that has completed path screening and structural pruning, high-response nodes and key connection structures are retained and reassembled to form a sparse network architecture, ultimately obtaining a simplified KAN network with structural readability and efficient reasoning capabilities.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] 1. This invention proposes an interpretable short-term wind power generation prediction method based on the Kolmogorov-Arnold network (KAN). Under multivariate input conditions, this method can efficiently model the nonlinear characteristics of new energy power generation by constructing a learnable network structure composed of one-dimensional B-spline functions. Furthermore, by introducing sparsification and symbolization mechanisms, it automatically extracts symbolic prediction expressions with clear physical meanings, thereby providing a structurally transparent and reliable foundation for power generation modeling, operation prediction, and control strategy design of new energy systems.

[0037] 2. This invention boasts advantages such as flexible structure and low training sample requirements, enabling rapid adaptation to various power generation scenarios in real-world conditions. By introducing the interpretability mechanism of the KAN network, it solves the problem of severe "black box" characteristics and lack of transparency in traditional deep learning models for wind power generation prediction. Compared with traditional neural networks, it significantly improves model readability and deployment convenience, providing a more practical and reliable technical solution for power generation modeling tasks.

[0038] 3. Compared with traditional deep learning models, the interpretable short-term power generation prediction method based on the Kolmogorov-Arnold network (KAN) in this invention has significant advantages in terms of strong interpretability and high generalization ability. It can automatically identify the nonlinear functional relationships between power generation characteristics and intuitively present the internal mechanism of the model in symbolic form. It realizes clear modeling of multivariate coupling relationships and transparent expression of model structure in new energy power generation scenarios. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating the implementation of an interpretable wind power generation prediction modeling method based on KAN network optimization in this embodiment.

[0041] Figure 2 This is a comparison chart of the training convergence of KAN and MLP models in this embodiment;

[0042] Figure 3 A comparison of the prediction performance of the KAN and MLP models in the wind power generation prediction task;

[0043] Figure 4 This is a visualization of the KAN model structure in this embodiment. Detailed Implementation

[0044] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Please see Figures 1-4This invention provides an interpretable short-term wind power generation prediction method based on the Kolmogorov-Arnold network, which includes the following steps:

[0046] S1: Collect multi-source data from the target power generation equipment over a historical timeframe to construct a predictive input feature set and obtain the power generation time series and its environmental correlation characteristics. Specifically, for any new energy power generation equipment, collect multi-source feature variables and power output data during its operation, preprocess and uniformly encode the raw data, and construct a predictive input-output dataset for KAN modeling, including:

[0047] S11: Historical Operational Data Acquisition: For the new energy power generation equipment to be modeled, historical operational data is collected under steady-state conditions during grid-connected operation. The collected data includes environmental characteristic parameters and temporal characteristic information. Simultaneously, the actual power generation value at the corresponding moment is recorded, forming a structured historical data sample set, providing fundamental support for subsequent power generation prediction modeling.

[0048] S12: Data Preprocessing and Standardization: All feature variables are preprocessed uniformly. First, the Z-Score standardization method is used to standardize the j-th feature value of the i-th sample. Transformed into:

[0049] ;

[0050] in, Let be the mean of a feature across all samples; Let $j$ be the standard deviation of the j-th feature across all samples. , The standardized feature values ​​have a mean of 0 and a variance of 1, which helps to accelerate model convergence and improve training stability.

[0051] Outliers are detected based on the 3σ principle and corrected using linear interpolation to ensure data continuity. Sine and cosine functions are used to encode the time feature, mapping the time variable t to... and The form is given by T, where T is the period. After processing, a standardized input feature vector is formed. and corresponding output target .

[0052] S13: Input-Output Sample Pair Construction: Based on the standardized time series data, sample pairs of input feature vectors and corresponding output target values ​​are constructed according to the time sequence. Specifically, a fixed-length historical observation sequence is set as the input feature, and input feature vectors formed at consecutive time points are extracted. The output target is the power generation data corresponding to a specified time step after the end of the selected input sequence. Following this method, the entire dataset is traversed, and N input-output sample pairs are constructed sequentially. This ultimately forms the training dataset for subsequent modeling. ,in Let represent the input variables for the i-th sample, such as ambient temperature, wind speed, etc. This indicates the corresponding power output.

[0053] S14: Dataset Partitioning and Formatting: The constructed dataset is divided into training, validation, and test sets in chronological order and converted into a data format acceptable to the KAN network. All input feature vectors must maintain dimensionality consistency and retain normalization parameters. Output power data is used to construct target prediction values. This ultimately forms a complete input-output data structure, providing a data foundation for subsequent KAN training.

[0054] S2: Based on the constructed input and output samples, build an initial KAN network and perform full training to obtain a basic model that fits the prediction relationship; specifically, based on the input and output samples, build an initial KAN network and train it to obtain a basic nonlinear mapping model for predicting the power generation of new energy equipment, including:

[0055] S21: Network Structure Initialization: Based on the Kolmogorov-Arnold theorem's multidimensional function decomposition theory, power generation prediction is modeled as a combinatorial optimization problem of multi-layer one-dimensional nonlinear functions. According to the core expression of the theorem:

[0056] ;

[0057] in It is a learnable one-dimensional function that connects the p-th node of the input layer to the q-th node of the first hidden layer. It is a learnable one-dimensional function that connects the q-th node of the last hidden layer to a single node in the output layer. This function is generalized to a multi-layer stacked KAN network with depth L and width sequence [n0, n1, ..., nL]. n0 is the input dimension, nL is the output layer, and each layer in the network... Both are passed through a function matrix F l To achieve nonlinear transformation:

[0058] ;

[0059] in To connect the first The i-th node to the i-th node of the layer The learnable one-dimensional function of the j-th node in the layer is derived from the B-spline basis functions. Combinatorial parameterization The coefficients w were initialized using a zero-mean Gaussian distribution. m1 Where G is the number of B-spline basis functions used. represents the weighted sum of all B-spline basis functions; m represents the index of the B-spline basis function.

[0060] S22: Network Parameter Training: Based on the structured training dataset D, construct the Kolmogorov-Arnold network prediction function and its parameter set θ to be optimized, defined as follows:

[0061]

[0062]

[0063] in, Let be the learnable transformation function for the m-th input feature. outer layer combined weight w k The set, Weights for inner affine transformation The set, Outer learnable function The parameter set; where k is the index of the outer summation, and K is the outer learnable function. The number of input features, M is the total number of input features.

[0064] Define the loss function with the objective of minimizing the mean squared error. ;

[0065] ;

[0066] in This indicates that the KAN network is applied to the i-th sample. The predicted value; Let represent the squared prediction error of the i-th sample.

[0067] The optimization problem is solved iteratively using the finite-memory quasi-Newton method (L-BFGS). Parameter updates follow ;

[0068] in The approximate inverse of the Hessian matrix for the current iteration step. This is the gradient of the loss function at the current parameter point. This is the optimal parameter set obtained after optimization; This indicates finding a loss function that makes the target defined. The smallest parameter θ. This represents the current parameter at the t-th iteration. This represents the updated parameters after the (t+1)th iteration.

[0069] Through the above iterative process, the spline coefficients and affine parameters in the KAN network can be efficiently optimized, and the nonlinear response characteristic function of the power generation of new energy equipment can be gradually fitted.

[0070] S23: Dynamic Raster Refinement: During training, the spline activation function within the KAN is refined using raster points at preset step intervals. Initially, the spline activation function... Based on the coarse mesh G1, its expression is as follows:

[0071] ;

[0072] in, Let i be the basis function of the i-th B-spline. for The order of c. i The corresponding learnable weights are given.

[0073] Subsequently, to address the complex response requirements within the activation region, a refined high-resolution mesh G2 (where G2 > G1) was introduced to reconstruct a more refined spline activation function. Its form of expression is:

[0074] ;

[0075] To maintain consistency between coarse and fine meshes, the fine mesh coefficient is optimized by minimizing the mean square error of both under the probability distribution p(x). :

[0076] ;

[0077] in This represents the expectation of a random variable x in probability distribution p(x).

[0078] This optimization process initializes high-resolution splines by minimizing the mean square deviation of the coarse and fine versions of the activation function, thereby effectively improving the KAN model's accuracy and generalization ability in characterizing the nonlinear characteristics of wind power generation without compromising its original fitting capability.

[0079] S24: Training Convergence and Error Monitoring. Monitor the changes in the training loss curve in real time to confirm that the model has reached a stable convergence state, and avoid overfitting or numerical divergence. If divergence occurs, roll back the optimization steps and adjust the learning rate or regularization parameters.

[0080] S3: Introducing L1 norm and entropy regularization constraints, the trained KAN network is pruned and sparsified to improve the sparsity of the network structure and the interpretability of features, achieving a sparse representation structure for the KAN network. This includes:

[0081] S31: Activation Function Response Statistics and Norm Constraint Construction: To improve the separability of responses from different activation paths in the KAN network, statistical modeling is performed on the response amplitudes of neurons in each layer. Let the... Activation output matrix of layer , , where N p For the sample size, R represents the number of active nodes in this layer, where R represents the real number space. It is an N p OK, For a real matrix of columns, its L1 norm is defined as:

[0082] ;

[0083] in Indicates the first The L1 norm of the activation output matrix of the layer. Indicates the first Layer activation output matrix In the diagram, the activation response value of the j-th neuron node corresponding to the i-th sample.

[0084] This norm is used to measure the average response strength of all nodes in the layer, serving as a penalty term to control structural complexity. By constraining the activation strength using the norm, the network can be induced to favor response expressions for the critical path.

[0085] S32: Construction of Normalized Entropy Regularization Term and Modeling of Distribution Balance. To further regulate the distribution of activation path responses among nodes, an entropy regularization mechanism is introduced. The normalized response of the j-th neuron node is defined as:

[0086] ;

[0087] in This represents the activation response of the j-th neuron node. The L1 norm of the activation response of the j-th neuron is represented. This represents the normalized response of the j-th neuron node. Indicates all of the previous layer The sum of the L1 norms of the nodes.

[0088] Based on this, the normalized information entropy of this layer is calculated as follows:

[0089] ;

[0090] in Indicates the first Normalized information entropy of the layer.

[0091] The smaller the entropy value, the more concentrated the activation response, which helps guide the network to form a focused sparse path structure; this term is added to the total loss as an auxiliary regularization quantity to improve the stability and interpretability of the structure.

[0092] S33: Construction of the Sparsified Objective Function and Full Training. Based on the above response intensity norm and normalized entropy term, the objective function including the structure regularization term is constructed as follows:

[0093] ;

[0094] Among them, L total Let L be the overall objective function of the KAN network. pred For the original predicted loss, , The weight hyperparameter for the regularization term represents the degree to which the network adjusts for sparsity and distribution. By optimizing this objective function, the response of non-critical neuron paths is suppressed during the training phase, laying the foundation for subsequent pruning operations and interpretable modeling.

[0095] S4: Based on the sparse training results, perform structural pruning on low-response nodes to construct a KAN network with readable structure. This includes:

[0096] S41: Activation Path Pruning and Structure Compression Based on Sparse Responses. After completing sparse training with L1 norm and entropy constraints, automatic pruning is performed on neuronal pathways with weak response amplitudes and low entropy distribution concentration. Specifically, for the i-th node, if its average activation response satisfies:

[0097] ;

[0098] It is then determined to be a low-contribution path, where θ1 is the response threshold, corresponding to redundant feature expression. This represents the activation response value of the i-th node corresponding to the k-th sample. Removing such nodes from the network connection graph and simultaneously updating the weight matrices of adjacent layers simplifies the structure and optimizes inference efficiency.

[0099] S42: KAN Network Reconstruction and Readable Structure Generation Based on Pruning Results. For the KAN network that has undergone path filtering and structural pruning, high-response nodes and key connection structures are retained and reassembled to form a sparse network architecture. The final result is a simplified KAN network with structural readability and efficient reasoning capabilities, which can provide effective support for the deployment and operational status analysis of wind power prediction.

[0100] The specific working principle of this invention:

[0101] First, multi-source historical data is collected from the target new energy power generation equipment to obtain an input feature set including environmental characteristic parameters and time characteristic information, and the actual power generation value at the corresponding time is recorded simultaneously. The raw data is preprocessed by the Z-Score standardization method to eliminate dimensional differences and correct outliers. Sine and cosine coding is used for time variables to characterize periodic characteristics. Finally, a unified and standardized input feature vector and power generation output sample pair are constructed.

[0102] Based on the Kolmogorov-Arnold theorem, an initial KAN network structure is built, and a nonlinear mapping model is constructed through a learnable B-spline activation function and affine transformation parameters. The loss function is optimized using a finite-memory quasi-Newton method, and the fitting accuracy is improved by combining a dynamic grid thinning mechanism, thus achieving efficient modeling between power generation and multivariate features.

[0103] By introducing L1 norm and entropy regularization constraints, the KAN network is trained to be sparse, suppressing low-contribution activation paths. By constructing an objective function that includes a structure regularization term, the network is guided to form a focused sparse path structure, laying the foundation for pruning.

[0104] Based on the sparse training results, low-response nodes are automatically pruned and redundant connections are removed to achieve structural compression. Then, key nodes and high-response paths are retained to form a structurally readable KAN network for deployment and operation analysis of wind power prediction.

[0105] To verify the effectiveness of the method, it was applied to a wind power generation prediction scenario. Figure 2 The model loss convergence process of KAN (red line) and MLP (blue line) is shown. KAN loss narrows rapidly and has a lower steady-state value. The comparison verifies its fast convergence and high accuracy in wind power prediction, which is suitable for the needs of the scenario. Figure 3 The performance of the MLP (top) and KAN (bottom) models in wind power prediction is presented separately. Actual power is represented by the solid blue line, while predicted power is represented by the dashed orange lines (MLP and KAN), respectively. Error ranges are indicated by gray filled areas. The comparison shows that the KAN model's prediction curve more closely matches the actual power trajectory, has a narrower error range, and exhibits more stable diurnal fluctuations, verifying its higher prediction accuracy and robustness in scenarios with fluctuating wind power. Figure 4 The multi-layer network topology is presented using nodes (black dots) and connections (gray lines). The upper layer diverges to the middle layer, and the middle layer crosses to the lower layer. The information processing architecture of the visualization analysis model is used to verify the rationality of the topology design for feature extraction and fusion, and to provide a structural basis for prediction performance.

[0106] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An interpretable short-term wind power generation prediction method based on Kolmogorov-Arnold networks, characterized in that, include: S1: Collect multi-source data from the target power generation equipment within a historical time range, construct a predictive input feature set, and obtain the power generation time series and its environmental correlation characteristics; S2: Based on the constructed input and output samples, build an initial KAN network and perform full training to obtain a basic model that fits the predicted relationship; S3: Introduce L1 norm and entropy regularization constraints to prune and sparsify the trained KAN network, thereby realizing the sparse representation structure of the KAN network. S4: Based on the sparse training results, structural pruning is performed on low-response nodes to construct a KAN network with structural readability.

2. The interpretable short-term wind power generation prediction method based on the Kolmogorov-Arnold network according to claim 1, characterized in that, In step S1, multi-source data is collected from the target power generation equipment within a historical timeframe. Specifically, for any target power generation equipment, multi-source feature variables and power output data during its operation are collected to obtain raw data. The raw data is then preprocessed and uniformly encoded to construct a prediction input-output dataset for KAN modeling, including: S11: Historical Operation Data Acquisition: For the target power generation equipment to be modeled, historical operation data is collected under steady-state conditions during grid-connected operation, including environmental characteristic parameters and time characteristic information; at the same time, the actual power generation value at the corresponding moment is recorded to form a structured historical data sample set; S12: Data Preprocessing and Standardization: All feature variables are preprocessed uniformly using the Z-Score standardization method, which standardizes the j-th feature value of the i-th sample. Transformed into standardized eigenvalues Based on the 3σ principle, outliers are detected and corrected using linear interpolation. Sine and cosine functions are used to encode the time features, mapping the time variable t to sin(2πt / T) and cos(2πt / T) forms, where T is the period. After processing, a standardized input feature vector is formed. and corresponding output target ; S13: Input-output sample pair construction: Based on the standardized time series data, construct sample pairs of input feature vectors and corresponding output target values ​​according to the time sequence. Specifically, set a fixed-length historical observation sequence as the input feature, and extract a series of consecutive time points to form the input feature vector. The output target is the power generation data corresponding to a specified time step after the end of the selected input sequence. Following this method, the entire dataset is traversed, and N input-output sample pairs are constructed sequentially. This ultimately forms the training dataset D for subsequent modeling. S14: Dataset Partitioning and Format Encapsulation: The constructed training dataset is divided into training, validation, and test sets in chronological order and converted into a data format acceptable to the KAN network. All input feature vectors must ensure dimensionality consistency and retain normalization parameters. Output power data is used to construct target prediction values, ultimately forming a complete input-output data structure.

3. The interpretable short-term wind power generation prediction method based on the Kolmogorov-Arnold network according to claim 2, characterized in that, In step S2, based on the constructed input and output samples, an initial KAN network is built and fully trained, including: S21: Network structure initialization: Based on the Kolmogorov-Arnold theorem's multidimensional function decomposition theory, power generation prediction is modeled as a combinatorial optimization problem of multi-layer one-dimensional nonlinear functions; S22: Network parameter training: Based on the training dataset D, construct the Kolmogorov-Arnold network prediction function and its parameter set θ to be optimized. Define the loss function with the goal of minimizing the mean square error, and use the finite memory quasi-Newton method to iteratively solve the optimization problem. S23: Dynamic Raster Refinement: During training, the spline activation function within the KAN is refined by raster points according to a preset step interval. In the initial stage, the activation function is constructed based on a coarse grid G1. Subsequently, to meet the complex response requirements within the activation region, a refined high-resolution grid G2 is introduced to reconstruct a finer spline activation function, where G2 > G1. S24: Training convergence and error monitoring: Monitor the changes in the training loss curve in real time to confirm that the model has reached a stable convergence state.

4. The interpretable short-term wind power generation prediction method based on the Kolmogorov-Arnold network according to claim 3, characterized in that, In step S3, the trained KAN network is pruned and sparsified, including: S31: Activation Function Response Statistics and Norm Constraint Construction: Statistical modeling of the response amplitudes of neurons in each layer is performed, assuming... The layer activation output matrix is , where N p For the sample size, Given the number of active nodes in this layer, the L1 norm is defined as: ； S32: Construction of Normalized Entropy Regularization Term and Modeling of Distribution Balance: To standardize the distribution of activation path responses among nodes, an entropy regularization mechanism is introduced, defining the normalized response of the j-th neuron node as: ； The normalized information entropy of this layer is calculated as follows: ； S33: Construction of the sparsified objective function and full training: Based on the response intensity norm and normalized entropy term in step S2, construct an objective function that includes a structure regularization term: ； in, For the original predicted loss, , The weight hyperparameter for the regularization term represents the degree to which the network adjusts for sparsity and distribution.

5. The interpretable short-term wind power generation prediction method based on the Kolmogorov-Arnold network according to claim 4, characterized in that, In step S4, structural pruning is performed on low-response nodes based on the sparsity training results, including: S41: Activation Path Pruning and Structure Compression Based on Sparse Responses: After completing sparse training with L1 norm and entropy constraints, automatic pruning is performed on neuronal pathways with weak response amplitudes and low entropy distribution concentration. Specifically, for the i-th node, if its average activation response satisfies: If θ1 is the response threshold, it is determined to be a low contribution path, corresponding to redundant feature expression. Such nodes are removed from the network connection graph, and the weight matrix of adjacent layers is updated synchronously. S42: KAN Network Reconstruction and Readable Structure Generation Based on Pruning Results: For the KAN network that has completed path filtering and structural pruning, high-response nodes and key connection structures are retained and reassembled to form a sparse network architecture, ultimately obtaining a simplified KAN network with structural readability and efficient reasoning capabilities.

Images