A neural network wind power prediction method based on a fusion mechanism

Abstract

The application relates to the field of wind power generation power prediction, and discloses a wind power generation power prediction method fusing a mechanism explainable neural network, which comprises the following steps: 1) data sample acquisition and preprocessing; 2) constructing a hierarchical feature extraction network based on a Kolmogorov-Arnold network; the hierarchical feature extraction network comprises an adaptive feature extraction submodule and a time sequence feature extraction submodule; 3) constructing a mechanism-fused Kolmogorov-Arnold prediction network; and 4) model training and outputting a prediction value; the application introduces a physical mechanism in a wind power generation process, guides a model training process, and improves the reliability and accuracy of prediction.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction, and more specifically to a wind power prediction method that incorporates a neural network with an interpretable mechanism. Background Technology

[0002] Wind energy is a key component of the global renewable energy system. In recent years, driven by both policy support and technological innovation, and based on the large-scale development of high-quality onshore and offshore wind energy resources, the wind power industry has achieved leapfrog development, demonstrating a positive role in building a safe and efficient energy system, addressing climate change, and promoting sustainable development. However, the inherent intermittency and power output fluctuations of wind power generation significantly increase the complexity and difficulty of grid security and stability operation. To address this core challenge, constructing a refined data acquisition system covering all time and space scales, and developing high-precision and robust wind power prediction (WPP) models and algorithms, are key technological pathways to improve wind power absorption efficiency and ensure the stable and efficient operation of the power system.

[0003] Currently, research on wind power generation prediction (WPP) methods is mainly divided into two major directions based on differences in modeling paradigms: mechanism-driven and data-driven. Mechanism-driven WPP methods construct numerical models based on physical mechanisms to characterize the conversion process from wind energy to electricity. Therefore, they heavily rely on high-resolution meteorological data and Numerical Weather Prediction (NWP) systems, resulting in a computationally intensive and time-consuming model solving process, making it difficult to meet the stringent timeliness requirements of real-time power generation prediction. Data-driven WPP methods, on the other hand, deeply mine historical operational data and meteorological and environmental physical parameters to construct end-to-end nonlinear mapping prediction models. They possess core advantages such as high efficiency in model training and prediction computation, strong ability to capture nonlinear relationships, and high prediction accuracy. Among these, deep learning methods, with their unique deep network architecture and powerful automatic feature extraction capabilities, have become one of the important research directions in the field of wind power generation prediction.

[0004] However, the distributed representation characteristics and multi-layer nonlinear transformations of deep learning obscure the causal link between input features and the final output, leading to a lack of interpretability in the model's internal decision-making mechanism and limiting its application potential in real-world production environments requiring high decision-making transparency. Therefore, constructing a neural network model architecture with interpretable parameters and forming a deep integration modeling framework based on domain mechanisms is a crucial technical path to improve the reliability of the WPP method in actual wind power generation processes. Against this backdrop, Kolmogorov-Arnold Networks (KAN) offer a novel paradigm for parameter-interpretable neural networks. Based on the Kolmogorov-Arnold representation theorem, this architecture achieves global approximation of complex nonlinear functionals through parameterized polynomial basis functions, quantifying the contribution of multidimensional input features to output prediction with closed-form expressions, thus providing mathematical transparency and interpretability to the decision-making process of the black-box model. Based on the aforementioned characteristics of the KAN model, this invention constructs a wind power prediction method that integrates mechanistic interpretable neural networks, incorporating the physical mechanisms of the wind power generation process to guide the model training process and improve the reliability and accuracy of predictions. Summary of the Invention

[0005] To address the issues of insufficient accuracy and parameter interpretability in existing deep learning models for wind power prediction, this invention proposes a wind power prediction method that integrates a neural network with interpretable mechanisms.

[0006] This method is based on the energy conservation mechanism model of wind power generation, uses the Kolmogorov-Arnold network to extract the spatiotemporal dynamic characteristics of input variables, and finally outputs the predicted power generation value under the guidance of the physical mechanism.

[0007] The specific technical solution is as follows:

[0008] A wind power prediction method that integrates mechanisms that can explain neural networks, the method comprising the following steps:

[0009] Step 1) Data sample acquisition and preprocessing: First, standardize the wind power dataset to unify the scale of each feature; then, extract continuous data subsets through a sliding time window to construct prediction samples;

[0010] Step 2) Construct a hierarchical feature extraction network based on the Kolmogorov-Arnold network: The hierarchical feature extraction network includes an adaptive feature extraction submodule and a temporal feature extraction submodule;

[0011] Step 2.1) The adaptive feature extraction submodule is built on the Kolmogorov-Arnold representation theorem, and integrates the basic linear activation transformation and the B-spline basis nonlinear transformation to achieve multi-dimensional static feature adaptive mapping;

[0012] Step 2.2) The temporal feature extraction submodule combines the Kolmogorov-Arnold network with a temporal cyclic gating structure to construct a temporal Kolmogorov-Arnold network. Through the gating mechanism, a B-spline high-order nonlinear mapping is introduced to construct temporal candidate states. The temporal cell states and hidden states are iteratively updated to mine the internal temporal correlations and high-dimensional potential features of wind power data and output multiple sets of separate temporal feature representations.

[0013] Step 3) Construct a mechanism-integrated Kolmogorov-Arnold prediction network:

[0014] Step 3.1) Establish a wind power physical mechanism model that includes wind energy capture and energy loss, and embed the physical mechanism constraints into the Kolmogorov-Arnold network;

[0015] Step 3.2) Based on the temporal representation output by the temporal feature extraction module and the original environmental features, conduct mechanism fusion modeling, construct energy conservation mechanism loss through numerical integration, combine power prediction loss to form a composite loss function, and generate high-dimensional feature representation of fused mechanism information under physical prior constraints.

[0016] Further, in step 1): obtain the wind power dataset, use the historical feature matrix with a duration of t and a feature dimension of d as the input sample, and predict the wind power at the next moment; use standardization transformation to unify the dimensions of each feature, and perform normalization conversion on the original features based on the mean and standard deviation of each feature, so that all features are converted to the same numerical scale.

[0017] Further, in step 2.1), an adaptive feature extraction module based on KAN is constructed: a KAN network is built based on the Kolmogorov-Arnold representation theorem. The KAN network performs a bi-branch parallel nonlinear mapping on the input vector of each layer. In the implementation of the KAN network, given the input vector of the Lth layer... KAN decomposes the mapping into two parts: a basic nonlinear transformation and a B-spline basis function expansion. Its output... The calculation form is shown in the following formula:

[0018]

[0019] in Based on the Sigmoid activation function, Describes the B-spline basis functions. and These are the corresponding weight matrices; by superimposing and fusing low-order basic mappings and high-order spline nonlinear mappings through a two-branch superposition, adaptive high-dimensional nonlinear extraction is completed for the standardized multi-dimensional original features of wind power.

[0020] Step 2.2) Constructing a KAN-based temporal feature extraction module: The KAN network is fused with a recurrent gating architecture to construct a temporal Kolmogorov-Arnold network (TKAN); at time steps... The TKAN basic unit receives the current input. Compared to the previous hidden state The gated variables are obtained through linear weighting. Then, the input gate is generated synchronously using the Sigmoid activation function. Forgotten Gate Output gate and candidate gate Candidate states with high-order nonlinearity are obtained using B-spline activation mapping. :

[0021]

[0022] in For the j-th B-spline function, and For learnable parameter matrix, This represents the dot product operation, N. b The hidden layer dimension is used; then, the temporal cell state is updated based on each gate and the candidate state. With hidden state Complete the extraction of time-series features from wind power data.

[0023] Further, step 3.1) involves constructing an energy conservation mechanism model:

[0024] Based on aerodynamic formulas, wind turbines capture wind energy, and combined with the ideal gas law under standard atmospheric pressure, ambient temperature is dynamically adjusted. Corresponding air density The cumulative wind energy during the time period t0 to t1 is obtained by integrating wind power over time.

[0025]

[0026]

[0027] in, Indicates wind power. Let A be the air density, A be the swept area of ​​the wind turbine rotor, v be the wind speed, and C be the air density. p This is the adjustment coefficient;

[0028] And construct the loss energy estimation equation:

[0029]

[0030] in, Represents the temperature of internal components of the wind turbine With ambient temperature The difference, Used to represent unknown heat loss coefficient Indicates average active power;

[0031] Based on the Kolmogorov-Arnold representation theorem, wind power The calculation formula can be converted into the following form:

[0032]

[0033] in, Standard atmospheric pressure The gas constant of air. Ambient temperature;

[0034] Based on the energy conservation law that wind power generation follows, an energy conservation mechanism model is established.

[0035]

[0036] in For the output electrical energy, Wind energy captured by wind turbines This refers to energy loss during the power generation process.

[0037] Step 3.2) Constructing a mechanism fusion model:

[0038] Coupled modeling of energy conservation mechanism and KAN network: The standardized historical wind power data X is divided into historical active power. and other environmental characteristic variables For the two types of data, the TKAN temporal feature extraction module was first used to extract the temporal features of each type of data. and Then, the features are concatenated together using linear projection to form the fused temporal feature f;

[0039]

[0040] in, and For learnable projection matrices, This indicates feature splicing.

[0041] Subsequently, combining the energy conservation mechanism model from step 3.1), a KAN wind turbine generation mechanism module is constructed, integrating the time-series characteristic f and environmental characteristic variables. After splicing, input into the KAN wind turbine generator theory module to predict active power. :

[0042]

[0043] Simpson's numerical integration method is used, based on historical active power. Calculate the actual cumulative electrical energy Constructing an energy conservation mechanism to address losses:

[0044]

[0045] Constructing the projection matrix Will Projecting onto a higher-dimensional space yields the mechanistic features. The mechanism characteristics are combined with the time-series characteristics of historical power. By splicing, the mechanistic and temporal fusion features are obtained. , will integrate features After being merged with historical data X, it is used as input. Construct the first TKAN module residual connection and output intermediate features. :

[0046]

[0047] Construct the second KAN module residual connection to obtain the active power. Predicted values:

[0048]

[0049] Construct the power prediction loss function:

[0050]

[0051] Introducing weight hyperparameters Balance loss calculation yields the total loss. l :

[0052]

[0053] Based on total loss L all The entire network's learnable parameters are updated iteratively in reverse to complete the few-sample wind power modeling that incorporates the physical mechanism of energy conservation.

[0054] The beneficial effects of this invention are as follows: By deeply integrating the Kolmogorov-Arnold Representation Theorem (KAN) with the physical mechanism of wind power energy conservation, this invention overcomes the shortcomings of traditional deep learning "black box" models, such as lack of interpretability and the tendency of purely data-driven models to violate physical laws. The method utilizes the closed-form expression of KAN and B-spline basis functions to quantify the contributions of multidimensional features, endowing the model's decision-making process with mathematical transparency. Simultaneously, it embeds the mechanistic constraints of wind energy capture and heat dissipation losses into the composite loss function, guiding and strongly regulating the optimization direction of network parameters. Combined with adaptive spatial extraction and the TKAN temporal gating mechanism, the model can accurately capture the spatiotemporal nonlinear correlations of wind power data. Case studies demonstrate that this invention significantly outperforms benchmark models such as LSTM and GRU in both NRMSE and NMAE metrics, maintaining extremely high prediction accuracy, strong robustness, and decision credibility even under limited sample conditions, which is of great significance for ensuring the stable operation of the power system. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a comparison chart of the prediction results of the method of this invention and the benchmark model;

[0057] Figure 3 This is an error box plot of the method of this invention and the benchmark model. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0059] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0060] Reference Figures 1-3 A wind power prediction method that integrates neural networks with a fusion mechanism, the method comprising the following steps:

[0061] (1) Data sample acquisition and preprocessing:

[0062] This invention uses a wind power generation dataset as a validation dataset, which was collected from a wind power plant in Zhejiang. The table below shows the composition of the feature variables of the dataset, and the rated power of the wind turbine is 2 kilowatts.

[0063] Table 1 Introduction to Wind Power Generation Datasets

[0064] Active power Kw wind speed m / s Rear bearing temperature ℃ Gear oil temperature ℃ Wheel hub temperature ℃ Ambient temperature ℃ Transformer battery temperature ℃

[0065] Step 1.1: Data Standardization. Given a wind power dataset, This represents the matrix of all historical feature variables at time step t, where d represents the dimension of the feature variables, and the prediction target is the power generation at time step t+1. The specific process for standardizing it is as follows:

[0066] (1)

[0067] in, and Let i and j represent the i-th feature before and after standardization, respectively. and Let represent the mean and standard deviation of the i-th feature, respectively.

[0068] (2) Construct a KAN-based spatiotemporal feature extraction module.

[0069] Step 2.1: Construct a KAN-based adaptive feature extraction module

[0070] KAN originates from the Kolmogorov-Arnold representation theorem, which states that any continuous multivariate function with n inputs... All of these can be precisely represented by a finite combination of continuous univariate functions in the following form:

[0071] (2)

[0072] in and It is a univariate continuous function. This conclusion provides a theoretical basis for high-dimensional function decomposition, enabling complex multidimensional mappings to be realized through several one-dimensional mappings and linear combinations. In the implementation of the KAN network, given the input vector of the Lth layer... KAN decomposes the mapping into two parts: a basic nonlinear transformation and a B-spline basis function expansion. Its output... The calculation form is shown in formula (3):

[0073] (3)

[0074] in Based on the Sigmoid activation function, Describes the B-spline basis functions. and These are the corresponding weight matrices. The B-spline basis functions are defined by the following recursive formula:

[0075] (4)

[0076] (5)

[0077] in The values ​​of the node vectors are represented by i and j, which are the coordinates of the sequence positions of the B-spline basis functions.

[0078] Step 2.2: Construct a KAN-based temporal feature extraction module

[0079] To meet the needs of time series modeling and capture high-dimensional latent space representations of complex temporal features, Temporal Kolmogorov-Arnold Networks (TKANs) introduce a recurrent structure, combining the nonlinear representation capabilities of the original KAN model with temporal dependency modeling. At time steps... The TKAN basic unit receives the current input. Compared to the previous hidden state The gating variable is calculated using the following formula. :

[0080] (6)

[0081] Where W and U are weight matrices, and b is the bias term. Its input gate... Forgotten Gate Output gate and candidate gate It is obtained through the following formula:

[0082] (7)

[0083] in The basic sigmoid activation function is used. Then, B-spline activation mapping is used to obtain candidate states with higher-order nonlinearities. The details are as follows:

[0084] (8)

[0085] in For the j-th B-spline function, and For learnable parameter matrix, This represents the dot product operation, N. b This represents the dimension of the hidden layer. Subsequently... Cellular state at any given moment With hidden layer state Update according to the gating mechanism:

[0086] (9)

[0087] (10)

[0088] in This is the activation function.

[0089] (3) Construct a mechanism-integrated Kolmogorov-Arnold prediction network.

[0090] Step 3.1: Construct an energy conservation mechanism model:

[0091] The process of wind power generation follows the following law of conservation of energy:

[0092] (11)

[0093] in For the output electrical energy, Wind energy captured by wind turbines This represents energy loss during the power generation process. The output electrical energy during the time period t0 to t1. From active power Obtained by time integration

[0094] (12)

[0095] in This represents the integral over time from time t0 to t1, i.e. P(s) represents the power generation at time s.

[0096] Wind power that captures wind energy Calculated using the following aerodynamic formula:

[0097] (13)

[0098] in Let A be the air density, A be the swept area of ​​the wind turbine rotor, v be the wind speed, and C be the air density. p is a coefficient.

[0099] air density The dynamic estimation can be performed using the ideal gas law under standard atmospheric pressure, specifically:

[0100] (14)

[0101] in Standard atmospheric pressure The gas constant of air. Let C be the ambient temperature. This method calculates the air density under ideal conditions; however, actual wind power generation is not entirely ideal, so deviations will be accounted for by the coefficient C. pAdjustments are made. Then, the cumulative wind energy during the time period t0 to t1 can be calculated by integrating the wind power over time.

[0102] (15)

[0103] Considering that some of the energy loss will eventually be dissipated into the environment as heat, the following energy loss estimation equation is constructed:

[0104] (16)

[0105] in Represents the temperature of internal components of the wind turbine With ambient temperature The difference, Used to represent unknown heat loss coefficient This represents the average active power. The above equations construct a mechanistic model of the wind power generation process.

[0106] Step 3.2: Constructing a mechanism fusion model:

[0107] The Kolmogorov-Arnold representation theorem described in step 2.1 states that any continuous multivariate function can be represented by a finite number of continuous univariate functions, and the wind power calculation formula in formula (13) can be transformed into the following form:

[0108] (17)

[0109] By comparing formulas (2) and (17), it is found that, given the input variables And v, construct a KAN network to predict P wind The process essentially involves finding the optimal parameters using a data-driven approach and calculating them according to theoretical formulas. Based on this, this paper employs a KAN network model built on the energy conservation mechanism. After inputting the original variables, the coefficients are automatically calculated during the modeling process. This ensures that the modeling process conforms to the physical mechanism and leverages the advantages of a data-driven approach, as detailed below:

[0110] Historical data X can be divided into two parts: historical active power. and other environmental characteristic variables First, the TKAN module is used to extract the temporal features of these two types of data respectively. and Where B is the batch size, S is the time window size, and H is the hidden layer dimension. Then, linear projection is used to concatenate the features together as the fused temporal feature f.

[0111] (18)

[0112] in and For learnable projection matrices, This indicates feature splicing.

[0113] Subsequently, combining the energy conservation model from step 3.1, a KAN-based wind turbine generator theory module is constructed, incorporating time-series characteristics f and environmental characteristic variables. Predicted active power Specifically:

[0114] (19)

[0115] Real accumulated electrical energy Using Simpson numerical integration, through historical active power Calculation yielded:

[0116] (20)

[0117] in For time step, The weights of Simpson's numerical integral, Indicates the t-th t The active power at the time step. At this point, the energy conservation mechanism loss of the following model can be obtained:

[0118] (twenty one)

[0119] By employing the above mechanism fusion modeling method, the nonlinear relationship between active power and environmental characteristic variables in historical data was captured. This will further improve the model's ability to predict target variables. This approach utilizes temporal and nonlinear KAN modules to construct dual residual connections, enhancing the training stability of deep networks and improving their ability to model complex power dynamics. To this end, a projection matrix is ​​first constructed. Will Projected into a high-dimensional space and used as a mechanistic feature. Then it was compared with the time-series characteristics of historical power. By splicing, the mechanistic and temporal fusion features are obtained. Then the fusion features will be used. After being merged with historical data X, it is used as input. And construct the first TKAN module residual connection to output intermediate features. :

[0120] (twenty two)

[0121] Subsequently, the residual connection of the second KAN module was constructed to obtain the active power. Predicted values:

[0122] (twenty three)

[0123] The final power prediction loss function is:

[0124] (twenty four)

[0125] The final total loss is:

[0126] (25)

[0127] in The weight hyperparameters are used for balancing loss calculation.

[0128] (4) Model performance evaluation: In order to evaluate the method of the present invention more objectively, the evaluation indexes root mean square error (RMSE) and mean absolute error (MAE) are introduced.

[0129] Step 4.1: Model Performance Evaluation Metrics

[0130] Normalized Mean Absolute Error (NMAE) and Normalized Root Mean Squared Error (NRMSE) are used to evaluate model performance.

[0131] (26)

[0132] (27)

[0133] in Indicates the true label, This represents the predicted value, where n represents the total number of samples. This indicates the rated power of the wind turbine.

[0134] Step 4.2: Model Training and Performance Metric Comparison

[0135] To compare the model performance, a Long Short-Term Memory (LSTM) network and a Gated Recurrent Unit (GRU) were used to compare with the proposed method. The dataset consisted of 900 data sets, divided into training and test sets in a 3:2 ratio. The model used an Adaptive Moment Estimation (Adam) optimizer to iteratively optimize the model parameters, with a learning rate of 0.0001 and 200 training epochs. The final model performance comparison structure is shown in Table 2 below.

[0136] Table 2 Model Performance Comparison

[0137] LSTM 0.1709 0.1498 GRU 0.1392 0.1195 Proposed 0.0931 0.0644

[0138] The results show that the proposed method has lower performance indicators than the LSTM and GRU methods, demonstrating the effectiveness and superiority of the proposed method. Figure 2 and Figure 3 Visualize the box plots of the model's prediction results and prediction errors separately. Figure 2 As can be seen, the prediction curve of the proposed method is closer to the true value, demonstrating the accuracy of the method. Figure 3 Box plots of prediction errors of the proposed method and traditional benchmark models (LSTM, GRU) on the test set are shown. Figure 3 As can be clearly seen, the error bin of the method of this invention is significantly flatter, and its median is closer to the zero error line, indicating that the overall prediction bias of the method of this invention is extremely small. At the same time, the upper and lower edges and outlier points of the method of this invention are significantly fewer and shorter than those of LSTM and GRU models. This fully demonstrates that under extreme conditions such as sudden changes in wind speed or unstable environments, the mechanism fusion model with physical prior constraints can exhibit extremely strong robustness and noise resistance, effectively avoiding large-scale prediction distortion.

[0139] The embodiments of this invention are validated using a real wind power generation dataset collected from a wind power plant in Zhejiang Province. Combined with the quantitative evaluation results in Table 2, the proposed method significantly surpasses existing technologies in all performance indicators: Specifically, the NRMSE of this method is only 0.0931, a 45.5% reduction compared to LSTM's 0.1709 and a 33.1% reduction compared to GRU's 0.1392; simultaneously, its NMAE is significantly reduced to 0.0644, a reduction of 57.0% and 46.1% compared to LSTM and GRU, respectively. The above comparison results based on real industrial data intuitively and clearly demonstrate that, under the same network training environment and with limited sample size, this invention deeply integrates the constraints of the physical mechanism of energy conservation, effectively avoiding the prediction distortion caused by the blind fitting of traditional pure data-driven black-box models. This allows the prediction curve to more closely match the actual power output trajectory, comprehensively improving the accuracy of wind power prediction while exhibiting extremely significant technological advancements.

[0140] This invention proposes a wind power prediction method that integrates a neural network with an interpretable mechanism. This method incorporates the energy conservation mechanism in the wind power generation process into data-driven modeling, effectively capturing the nonlinear relationships between data variables and guiding the model training process to conform to the process mechanism. Case studies comparing the method with multiple baseline models further demonstrate that it improves prediction accuracy and possesses universality and applicability.

[0141] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A wind power prediction method that integrates a neural network with an interpretable mechanism, the method comprising the following steps: Step 1) Data sample acquisition and preprocessing: First, standardize the wind power dataset to unify the scale of each feature; Then, a subset of continuous data is extracted using a sliding time window to construct a prediction sample; Step 2) Construct a hierarchical feature extraction network based on the Kolmogorov-Arnold network: The hierarchical feature extraction network includes an adaptive feature extraction submodule and a temporal feature extraction submodule; Step 2.1) The adaptive feature extraction submodule is built on the Kolmogorov-Arnold representation theorem, and integrates the basic linear activation transformation and the B-spline basis nonlinear transformation to achieve multi-dimensional static feature adaptive mapping; Step 2.2) The temporal feature extraction submodule combines the Kolmogorov-Arnold network with a temporal cyclic gating structure to construct a temporal Kolmogorov-Arnold network. Through the gating mechanism, a B-spline high-order nonlinear mapping is introduced to construct temporal candidate states. The temporal cell states and hidden states are iteratively updated to mine the internal temporal correlations and high-dimensional potential features of wind power data and output multiple sets of separate temporal feature representations. Step 3) Construct a mechanism-integrated Kolmogorov-Arnold prediction network: Step 3.1) Establish a wind power physical mechanism model that includes wind energy capture and energy loss, and embed the physical mechanism constraints into the Kolmogorov-Arnold network; Step 3.2) Based on the temporal representation output by the temporal feature extraction module and the original environmental features, conduct mechanism fusion modeling, construct energy conservation mechanism loss through numerical integration, combine power prediction loss to form a composite loss function, and generate high-dimensional feature representation of fused mechanism information under physical prior constraints.

2. The method for predicting wind power using a few-sample model based on fusion mechanism transfer modeling as described in claim 1, characterized in that, Step 1): Obtain the wind power dataset, use the historical feature matrix with a duration of t and a feature dimension of d as the input sample, and predict the wind power at the next moment; use standardization transformation to unify the dimensions of each feature, and perform normalization conversion on the original features based on the mean and standard deviation of each feature, so that all features are converted to the same numerical scale.

3. The method for predicting wind power using a few-sample model based on fusion mechanism transfer modeling as described in claim 1, characterized in that, Step 2.1) Constructing a KAN-based adaptive feature extraction module: A KAN network is built based on the Kolmogorov-Arnold representation theorem. The KAN network performs a bi-branch parallel nonlinear mapping on the input vector of each layer. In the implementation of the KAN network, given the input vector of the Lth layer... KAN decomposes the mapping into two parts: a basic nonlinear transformation and a B-spline basis function expansion. Its output... The calculation form is shown in the following formula: ； in Based on the Sigmoid activation function, Describes the B-spline basis functions. and These are the corresponding weight matrices; by superimposing and fusing low-order basic mappings and high-order spline nonlinear mappings through a two-branch superposition, adaptive high-dimensional nonlinear extraction is completed for the standardized multi-dimensional original features of wind power.

4. The method for predicting wind power using a few-sample model based on fusion mechanism transfer modeling as described in claim 1, characterized in that, Step 2.2) Constructing a KAN-based temporal feature extraction module: The KAN network is fused with a recurrent gating architecture to construct a temporal Kolmogorov-Arnold network (TKAN); at time steps... The TKAN basic unit receives the current input. Compared to the previous hidden state The gated variables are obtained through linear weighting. Then, the input gate is generated synchronously using the Sigmoid activation function. Forgotten Gate Output gate and candidate gate Candidate states with high-order nonlinearity are obtained using B-spline activation mapping. : ； in For the j-th B-spline function, and For learnable parameter matrix, This represents the dot product operation, N. b The hidden layer dimension is used; then, the temporal cell state is updated based on each gate and the candidate state. With hidden state Complete the extraction of time-series features from wind power data.

5. The method for predicting wind power using a few-sample model based on fusion mechanism transfer modeling as described in claim 1, characterized in that, Step 3.1) Constructing an energy conservation mechanism model: Based on aerodynamic formulas, wind turbines capture wind energy, and combined with the ideal gas law under standard atmospheric pressure, ambient temperature is dynamically adjusted. Corresponding air density The cumulative wind energy during the time period t0 to t1 is obtained by integrating wind power over time. ； ； in, Indicates wind power. Let A be the air density, A be the swept area of ​​the wind turbine rotor, v be the wind speed, and C be the air density. p This is the adjustment coefficient; And construct the loss energy estimation equation: ； in, Represents the temperature of internal components of the wind turbine With ambient temperature The difference, Used to represent unknown heat loss coefficient Indicates average active power; Based on the Kolmogorov-Arnold representation theorem, wind power The calculation formula can be converted into the following form: ； in, Standard atmospheric pressure The gas constant of air. Ambient temperature; Based on the energy conservation law that wind power generation follows, an energy conservation mechanism model is established. ； in For the output electrical energy, Wind energy captured by wind turbines This refers to energy loss during the power generation process.

6. The method for predicting wind power using a few-sample model based on fusion mechanism transfer modeling as described in claim 5, characterized in that, Step 3.2) Constructing a mechanism fusion model: Coupled modeling of energy conservation mechanism and KAN network: The standardized historical wind power data X is divided into historical active power. and other environmental characteristic variables For the two types of data, the TKAN temporal feature extraction module was first used to extract the temporal features of each type of data. and Then, the features are concatenated together using linear projection to form the fused temporal feature f; ； in, and For learnable projection matrices, Indicates feature splicing; Subsequently, combining the energy conservation mechanism model from step 3.1), a KAN wind turbine generation mechanism module is constructed, integrating the time-series characteristic f and environmental characteristic variables. After splicing, input into the KAN wind turbine generator theory module to predict active power. : ； Simpson's numerical integration method is used, based on historical active power. Calculate the actual cumulative electrical energy Constructing an energy conservation mechanism to address losses: ； Constructing the projection matrix Will Projecting onto a higher-dimensional space yields the mechanistic features. The mechanism characteristics are combined with the time-series characteristics of historical power. By splicing, the mechanistic and temporal fusion features are obtained. , will integrate features After being merged with historical data X, it is used as input. Construct the first TKAN module residual connection and output intermediate features. : ； Construct the second KAN module residual connection to obtain the active power. Predicted values: ； Construct the power prediction loss function: ； Introducing weight hyperparameters Balance loss calculation yields the total loss. l : ； Based on total loss L all The entire network's learnable parameters are updated iteratively in reverse to complete the few-sample wind power modeling that incorporates the physical mechanism of energy conservation.

Images