A few-shot wind power forecasting method based on fusion mechanism migration modeling

Abstract

The present application relates to the field of short-term wind power generation prediction, and discloses a few-sample wind power prediction method fusing mechanism migration modeling, step 1) wind power generation data preprocessing and dataset division; step 2) MIC characteristic variable screening; step 3) defining characteristic variables and labels on the source domain and the target domain; step 4) establishing a KAN mechanism fusion model on the source domain; step 5) source domain pre-training and physical guided migration to the target domain; step 6) model training, prediction and model performance evaluation on the target domain; the present application introduces a migration strategy of freezing the physical layer, migrates the trained physical network layer in the source domain to the target domain, and only fine-tunes the prediction module, so as to realize wind power modeling under the condition of low samples; the strategy fully retains the physical feature expression ability in the source domain, significantly improves the prediction accuracy and stability under the condition of small sample learning of the target domain, and shows good migration generalization ability.

Description

Technical Field

[0001] This invention relates to the field of short-term wind power generation prediction, specifically to a few-sample wind power prediction method that integrates mechanism transfer modeling. Background Technology

[0002] In power systems, high-precision prediction of wind power generation is a key technological aspect for ensuring the safe, stable, and economical operation of the power grid. With the increasing penetration of renewable energy sources such as wind power in the power structure, the uncertainty and volatility of wind power generation have significantly intensified, posing severe challenges to power dispatch and operation control. Traditional methods based on empirical modeling or physical simulation are ill-suited to the complex and ever-changing operating environment of wind farms. In recent years, data-driven models, represented by deep neural networks, have gained widespread attention in wind power prediction tasks. Their ability to learn complex time-series patterns from large-scale historical wind speed, direction, and power data has significantly improved the prediction accuracy and automatic modeling capabilities of these models.

[0003] However, deep learning models still face many challenges in practical deployment, including high sensitivity to data quality, neglect of potential physical constraints, and insufficient interpretability and generalization ability. Especially in wind power modeling, failing to consider the physical consistency of wind farm operation can lead to severe performance degradation in new environments. To address this, a physics-guided neural network approach has been proposed to integrate physical laws with the advantages of data-driven approaches, enhancing the model's credibility and robustness. This method significantly improves the model's ability to suppress non-physical anomalies and enhances the physical consistency and interpretability of prediction results by embedding constraints such as energy conservation.

[0004] Building upon this foundation, to further enhance the model's ability to process input features and express structural characteristics, this paper proposes a physical information modeling framework that integrates the Maximum Information Coefficient (MIC) and the feature extraction mechanism of a Kolmogorov-Arnold Network (KAN) based on the Kolmogorov representation theorem. MIC is used to select high-quality features most relevant to the target variable, avoiding interference from redundant or perturbed features in the modeling process. The KAN module, leveraging its powerful nonlinear modeling and feature decoupling capabilities, mines the physical structural information within the input variables, thereby constructing a more robust feature representation.

[0005] Furthermore, given that wind power projects in complex or harsh environments, such as offshore and high-altitude wind farms, often lack sufficient historical operational data due to short construction periods, initial equipment instability, or limited monitoring conditions, this invention introduces a physical layer freezing transfer strategy. This strategy transfers the trained physical network layer from the source domain to the target domain, fine-tuning only the prediction module to achieve wind power modeling under low-sample conditions. This strategy fully preserves the physical feature representation capabilities of the source domain, significantly improving prediction accuracy and stability under low-sample learning in the target domain, and demonstrating good transfer generalization ability. Overall, the unified framework integrating feature selection, physical modeling, and transfer learning constructed in this study provides a more scientific, interpretable, and engineering-practical solution for wind power prediction tasks. Summary of the Invention

[0006] To address the lack of short-term trend information and interpretability in wind power prediction using deep learning methods, this invention proposes a few-sample wind power prediction method that integrates mechanism transfer modeling. Since wind power generation involves the mutual conversion of various energies, this invention uses the law of conservation of energy as the physical knowledge in the wind power generation process and designs a residual constraint method to guide the model behavior to follow the underlying physical knowledge.

[0007] The technical solution adopted by this invention to solve its technical problem is:

[0008] A few-sample wind power prediction method integrating mechanism transfer modeling includes the following steps:

[0009] Step (1) Obtain the original sample data and filter out the physical auxiliary variables.

[0010] Step (2) Preprocessing of raw sample data and partitioning of dataset: The raw sample data is time-seriested using a moving window, and then the time-seriesd sample data is divided into training set and test set; In order to speed up the convergence of the model and reduce the training time, the training set data and test set data are normalized.

[0011] Step (3) Domain definition and prediction target: Obtain time-seriesed source and target domain data.

[0012] Step (4) KAN mechanism fusion model:

[0013] A simplified energy conservation equation is established. The measured power is concatenated with the target power using a recurrent neural network (RNN) model. The concatenated data is then used to calculate various energy parameters in the wind power generation process, and the calculation results are corrected using compensation coefficients. A KAN network based on the Kolmogorov representation theorem is introduced. The selected auxiliary variables are nonlinearly projected dimension by dimension to obtain new auxiliary features. The residual loss is calculated using the constructed simplified energy conservation equation, and the model parameters are optimized on the training set.

[0014] Step (5) Source Domain Pre-training and Physics-guided Transfer: Save the pre-trained model trained on the source domain, freeze the physical model and transfer it to the target domain. Use the adversarial training mode to optimize the pseudo-labels generated on the target domain, and fine-tune the training to optimize the model parameters. Finally, test it on the test set of the target domain.

[0015] Furthermore, the process of step (3) is as follows:

[0016] The source domain dataset is Include Historical data at each time step, including Let F be the auxiliary variables at time k. This represents the power generation at time k. The objective of the source domain is to utilize... Auxiliary variables at each time step Measure the power generation at the last time step For those containing time step The target domain data, the task is to utilize historical auxiliary variables. predict Power generation at any time The source domain model aims to capture the temporal dynamics of historical sequences and the nonlinear mapping relationship between auxiliary variables and target variables, while the target domain task is to predict the power generation at the next time step based on historical data.

[0017] Furthermore, the process of step (4) is as follows:

[0018] Step 4.1: Construct a simplified energy conservation equation:

[0019] Based on the energy conversion principle of wind turbines, a simplified energy conservation relationship is established:

[0020] Based on existing research and modeling methods based on the principle of energy conservation, a simplified energy conservation model for wind turbine operation was constructed by examining the conversion relationships between wind energy, mechanical energy, electrical energy, and thermal energy.

[0021] During the operation of a wind turbine, wind energy is converted into mechanical energy, electrical energy, thermal energy, and other forms of energy. However, considering that it is impossible to obtain all the data in the actual process to construct a complete energy conservation equation, this invention simplifies the energy conservation equation.

[0022] In the task of predicting wind power generation, the input data is processed through a sliding window to obtain a three-dimensional matrix. Where B represents the batch size. Indicates the time step. Indicates the dimension of the feature variable.

[0023] The mechanism model of wind power generation is based on the core principle of converting air kinetic energy into electrical energy, and the air kinetic energy can be expressed as: ;

[0024] in Let A be the air density and A be the swept area of ​​the wind turbine rotor. for Wind speed at any moment, C p This is the power coefficient. The formula reveals the nonlinear nature of the cubic relationship between power and wind speed. Furthermore, environmental factors during wind power generation will affect the power coefficient by influencing air density, thus affecting the final output power.

[0025] Inputting mechanical energy causes a coil to cut magnetic field lines in a magnetic field, generating electrical energy. During this process, a phase difference exists between voltage and current in alternating current, causing the current to lag behind the voltage. When the current and voltage are out of sync, the circuit contains not only active power capable of performing actual work, but also a portion of energy exchanged periodically between the power source and the load, known as reactive power.

[0026] Because capacitors absorb the lagging reactive power generated by inductance, they can be used... The power value in the capacitor at any given time is used to approximate the reactive power. ,use The power value measured at time represents Active power at any time .

[0027] During wind turbine power generation, the heat energy mainly consists of two parts: heat absorbed by the wind turbine itself and heat dissipated by the wind turbine. The absorbed heat energy is expressed as the product of mass, specific heat capacity, and temperature difference: Where Q is the absorbed heat energy, m is the mass of the object, and c is the specific heat capacity of the object. For an object in time Temperature size, This represents the initial temperature of the object.

[0028] Because there is strong convection between the air and the wind turbine during operation, and the temperature difference between the wind turbine and the air is small, radiative heat transfer is ignored, and only convective heat transfer between the wind turbine and the air is considered. The formula for convective heat transfer per unit time is: ;

[0029] Where e qex The convective heat transfer per unit time. t is the convective heat transfer coefficient. wall The temperature of the wind turbine wall. for Air temperature at time A ex Let T be the heat exchange area between the wind turbine and the air. Then, the total energy over time period T can be expressed as the energy per unit time over time T. Using the law of conservation of energy, we can obtain the energy conversion relationships: ;

[0030] The energy matrix W obtained over time T is expressed as: ;

[0031] Considering the losses during energy conversion, a compensation coefficient matrix is ​​set. To approximate the energy conservation equation, where C w To compensate for the loss of wind energy converted into mechanical energy, C is the compensation coefficient. r C is the compensation coefficient for capacitance conversion. qin For and C qex Let be the energy conversion compensation coefficients for heat absorption and heat release, respectively. Assuming that the compensation coefficients remain constant within the time period T, when there is sufficient data, a system of equations can be constructed using Newton's Cotes formula, and then the corresponding compensation coefficients can be solved using the method of undetermined coefficients. ;

[0032] in Let W be the inverse matrix of W.

[0033] To make the physical model more closely related to the neural network model, the original input defined in step 3 will be... The measurement value at the last time step is obtained by inputting it into the RNN model. And with the target power The specific method for splicing is as follows: ;

[0034] And recalculated ,in The energy matrix is ​​obtained by calculating using the spliced ​​data, and the final predicted electrical energy is:

[0035]

[0036] in The electrical energy is calculated using auxiliary variables and weighting functions in the physical model.

[0037] However, due to the difficulty in fitting variables to the target electrical energy in wind power data, the weight function obtained by simply inverting the undetermined coefficients method cannot reconstruct accurate values ​​well. To construct a power generation prediction model based on the mechanism model of wind power generation, this application uses a KAN network to reconstruct the physical weight function. This theorem states that any continuous multivariate function containing n inputs can be decomposed into a linear combination of finite univariate functions. Therefore, this application uses a weighted KAN network to approximate the mechanism model of wind power generation. ;

[0038] in This represents the mechanism function of wind power generation fitted by the weighted KAN network model. and This represents a trainable inner unary function. For trainable outer combination functions, Since represents the p-th feature variable, q represents the q-th univariate function. Because KAN decomposes complex multivariate functions into linear combinations of several univariate functions, it avoids the coupling of complex variables. Inner function and The outer function is able to focus on capturing the physical relationship between each input feature variable and the output variable. Its role is to integrate the contributions of multiple features, providing a basis for obtaining the expected physical output.

[0039] Wind power generation exhibits strong temporal dynamics. While the above mechanism data fusion modeling process extracts the nonlinear relationship between input and output, it does not consider these temporal dynamics. Given that energy is the accumulation of power over time, therefore, in terms of time... The energy within the time frame can be obtained by integrating power over time. Therefore, an average optimization method based on numerical integration is proposed, which first integrates the power over time. The power data within the time frame is numerically integrated to fuse the power time-time dynamic characteristics from multiple time steps into energy characteristics. Since the data consists of discrete values ​​sampled at fixed time intervals, the numerical integral is solved using the following Newton-Cotes formula to approximate the integral term in the above formula. The Newton-Cotes formula is shown below. ;

[0040] Where a and b are the endpoints of the region to be multiplied, the region to be multiplied will be divided into H equal parts. for For any feature variable in the equation, the k-th division point has numerical value. , These are the corresponding coefficients of Newton-Cotes' formula. Let be the function that needs to be multiplied.

[0041] The electrical energy E can then be calculated using the following formula: ;

[0042] Where E represents the accumulated electrical energy within the time period T. express The active power of a time step, where H represents the number of time steps within the time period. These are the constant coefficients of the corresponding Newton-Cotes formula.

[0043] To obtain the predicted value of accumulated electrical energy Design a KAN feature processing model to process the input sequence The physical feature matrix is ​​obtained after processing by the inner function of KAN. Then, the same numerical integration method is applied to each variable of the input sequence to obtain the integral feature matrix. This is used to characterize the time-cumulative properties of historical data. The physical feature matrix and the integral feature matrix are then concatenated and input into the outer function to obtain the predicted value of the accumulated electrical energy. .

[0044] Finally, to ensure physical consistency, the energy residual loss is defined as follows: ;

[0045] Where N represents the number of historical data intervals. and Let represent the predicted and actual electrical energy values ​​at time k, respectively.

[0046] Furthermore, the process of step (5) is as follows:

[0047] Step 5.1: Source Domain Model Training:

[0048] During the source domain model training phase, to meet the requirements of temporal dynamic modeling, an RNN model was used to extract the spatiotemporal features of wind power data. For the input historical auxiliary variable sequence... At each time step, the RNN updates the hidden state h using the following recursive relationship. t : ;

[0049] in, For input, W xhW hh These are the input and state transition weight matrices, respectively, b h For bias terms, This represents the tanh activation function. The final model output is... Through the following calculations: ;

[0050] Among them W hy To output the weight matrix, b y For output bias terms, This represents the output mapping function of the fully connected layer.

[0051] The RNN model described above implements the mapping relationship from the input sequence to the output sequence. To further train the model and make the prediction results closer to the true values, the following loss function is introduced to constrain the prediction error: ;

[0052] Then, the predicted output of the RNN model is... with input history auxiliary variable sequence After splicing, input the KAN mechanism model defined in step 4 and obtain the energy residual R. S Finally, the loss function for training the source domain model is: ;

[0053] in These are the hyperparameters used in the balancing model.

[0054] Step 5.2: Freeze the migration and fine-tune the target domain model:

[0055] Due to significant differences in geographical environment, meteorological conditions, and operational characteristics among different wind farms, directly applying the source domain model to the target domain can lead to performance degradation in prediction. To reduce cross-domain distribution differences, this study constructs a target domain model based on the frozen physical migration method. Historical data for the target domain... First, the RNN module is based on historical auxiliary variables. Predicted value of output active power In the source domain task, the RNN module aims to capture the nonlinear mapping relationship between variables and does not predict the power generation at the next time step. Therefore, during the transfer process, to retain the general features learned in the source domain model, this study uses parameter transfer to freeze most of the network parameters and only retrains some transferable layers on the target domain data to achieve domain adaptation of the model. Fine-tuning is performed on the fully connected layers in the transferred RNN module, and the loss function is constructed using the target domain data. ;

[0056] in , These are the predicted value and the actual power output by the RNN module at time step j in the target domain, respectively. Then, the predicted output of the RNN model... with input history auxiliary variable sequence After splicing, the data is input into the KAN mechanism model, and the energy residual R is obtained. T Finally, the loss function for training the source domain model is: ;

[0057] in These are the hyperparameters used to balance the model. Then, the RNN module and the KAN mechanism module are fixed, and the prediction output of the RNN model is... with input history auxiliary variable sequence After concatenation, the data is input into the KAN mechanism model to obtain pseudo-labels. However, this pseudo-label cannot accurately adapt to the mapping relationship between features. Therefore, to enhance domain adaptation, an adversarial mechanism is constructed and optimized using the following loss function: ;

[0058] It is the predicted power after training in the source domain and transferring it to the target domain. These are pseudo-labels generated by the physical model on the target domain. The discriminator model is represented. For the KAN mechanism module, it is updated using the following loss function to make it indistinguishable from the source domain from the discriminator's perspective, ultimately achieving a domain-invariant representation.

[0059] ;

[0060] The optimization goal of overall adversarial training is to ;

[0061] in To balance the hyperparameters of adversarial training, the KAN mechanism module and discriminator are updated alternately to achieve real-time correction of pseudo-labels in the target domain, thereby enabling accurate adversarial transfer.

[0062] The beneficial effects of this invention are as follows: The few-sample wind power prediction method based on fusion mechanism transfer modeling provided by this invention effectively overcomes the problem of data scarcity. Based on the pre-trained model in the source domain, it successfully transfers knowledge to the target domain and generates pseudo-labels through physical-guided transfer learning. Especially when data in the target domain is scarce and unlabeled, it still maintains high prediction accuracy. Compared with existing technologies, this invention not only optimizes the accuracy of wind power prediction but also enhances cross-domain adaptability and reduces the model's dependence on labeled data, demonstrating significant innovation and practical application value. Attached Figure Description

[0063] Figure 1 This is a flowchart of the novel method of the present invention;

[0064] Figure 2 This is a comparison diagram of the effects of the novel invention;

[0065] Figure 3 This is a novel power scatter plot of the present invention;

[0066] Figure 4 This is a scatter plot of the UDA predicted power of the present invention. Detailed Implementation

[0067] 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.

[0068] 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.

[0069] A method for predicting wind power using a fusion mechanism transfer model with few samples mainly includes: step 1) wind power data preprocessing and dataset partitioning; step 2) MIC feature variable screening; step 3) defining feature variables and labels in the source domain and target domain; step 4) establishing a KAN mechanism fusion model in the source domain; step 5) pre-training in the source domain and physically guiding the transfer to the target domain; (6) training the model in the target domain, predicting and evaluating the model performance.

[0070] Example: Taking the construction of active power prediction for wind power plants as an example, the following steps are included:

[0071] (1) Obtaining wind power generation data: Ten process variables were collected, namely, ambient temperature of the wind turbine generator, bearing shaft temperature, gearbox bearing temperature, gearbox oil temperature, generator winding 1 temperature, generator winding 2 temperature, hub temperature, main unit casing temperature, reactive power, and wind speed. The final predicted variable was active power. In order to improve the responsiveness of physical loss to key features and reduce the noise interference caused by some physical features, a feature selection method based on MIC was adopted to evaluate the statistical dependence between the energy features extracted from the original samples and the power generation. Considering that the physical model needs no less than 5 features to calculate the energy values ​​in the form of wind energy, thermal energy, etc., the top 5 features with the most information were finally selected for physical residual constraints. The features selected in the physical model were: gearbox bearing temperature, generator winding 1 temperature, generator winding 2 temperature, wind speed, and reactive power.

[0072] (2) Wind power data preprocessing and dataset partitioning: The original sample data will be time-seriesd using a moving window, and then the time-seriesd sample data will be divided into training set and test set; In order to speed up the model convergence and reduce the model training time, the training set data and test set data will be normalized.

[0073] Step 2.1: Split the dataset

[0074] For the source domain, the 4584 collected wind power data points were divided into two parts: 3667 samples were used as the training set, and the remaining 917 samples were used as the test set. For the target domain, the 4584 collected wind power data points were divided into two parts: 458 samples were used as the training set, and the remaining 4126 samples were used as the test set.

[0075] Step 2.2: Data Normalization Process

[0076] To accelerate model convergence and reduce training time, the data is normalized to reduce the differences between data units.

[0077] (3) Domain definition and prediction goal: Obtain time-seriesed source and target domain data.

[0078] The source domain dataset is Where S represents the source domain, containing Historical data at each time step, including Let F be the auxiliary variables at time k. This represents the power generation at time k. The objective of the source domain is to utilize... Auxiliary variables at each time step Measure the power generation at the last time step For those containing time step Target domain data, where T represents the target domain. Let F be the auxiliary variables in the source domain at time j. This represents the power generation in the target domain at time j, and its task is to utilize historical auxiliary variables. predict Power generation at any time The source domain model aims to capture the temporal dynamics of historical sequences and the nonlinear mapping relationship between auxiliary variables and target variables, while the target domain task is to predict the power generation at the next time step based on historical data.

[0079] (4) KAN mechanism fusion model: A simplified energy conservation equation is established, and the measured power is spliced ​​with the target power through a recurrent neural network (RNN) model. The spliced ​​data is then used to calculate various energy items in the wind power generation process, and the calculation results are corrected using compensation coefficients. A KAN network based on the Kolmogorov representation theorem is introduced, and new auxiliary features are obtained by nonlinearly projecting the selected auxiliary variables one dimension at a time. The residual loss is calculated using the constructed simplified energy conservation equation, and the model parameters are optimized on the training set.

[0080] Step 4.1: Construct a simplified energy conservation equation:

[0081] Based on the energy conversion principle of wind turbines, a simplified energy conservation relationship is established:

[0082] Based on existing research and modeling methods based on the principle of energy conservation, a simplified energy conservation model for wind turbine operation was constructed by examining the conversion relationships between wind energy, mechanical energy, electrical energy, and thermal energy.

[0083] During the operation of a wind turbine, wind energy is converted into mechanical energy, electrical energy, thermal energy, and other forms of energy. However, considering that it is impossible to obtain all the data in the actual process to construct a complete energy conservation equation, this invention simplifies the energy conservation equation.

[0084] In the task of predicting wind power generation, the input data is processed through a sliding window to obtain a three-dimensional matrix. Where B represents the batch size. Indicates the time step. Indicates the dimension of the feature variable.

[0085] The mechanism model of wind power generation is based on the core principle of converting air kinetic energy into electrical energy, where air kinetic energy can be expressed as... ;

[0086] in Let A be the air density and A be the swept area of ​​the wind turbine rotor. for Wind speed at any moment, C pThis is the power coefficient. The formula reveals the nonlinear nature of the cubic relationship between power and wind speed. Furthermore, environmental factors during wind power generation will affect the power coefficient by influencing air density, thus affecting the final output power.

[0087] Inputting mechanical energy causes a coil to cut magnetic field lines in a magnetic field, generating electrical energy. During this process, a phase difference exists between voltage and current in alternating current, causing the current to lag behind the voltage. When the current and voltage are out of sync, the circuit contains not only active power capable of performing actual work, but also a portion of energy exchanged periodically between the power source and the load, known as reactive power.

[0088] Because capacitors absorb the lagging reactive power generated by inductance, they can be used... The power value in the capacitor at any given time is used to approximate the reactive power. ,use The power value measured at time represents Active power at any time .

[0089] During wind turbine power generation, the heat energy mainly consists of two parts: heat absorbed by the wind turbine itself and heat dissipated by the wind turbine. The absorbed heat energy is expressed as the product of mass, specific heat capacity, and temperature difference: Where Q is the absorbed heat energy, m is the mass of the object, and c is the specific heat capacity of the object. For an object in time Temperature size, This represents the initial temperature of the object.

[0090] Because there is strong convection between the air and the wind turbine during operation, and the temperature difference between the wind turbine and the air is small, radiative heat transfer is ignored, and only convective heat transfer between the wind turbine and the air is considered. The formula for convective heat transfer per unit time is: ;

[0091] Where e qex The convective heat transfer per unit time. t is the convective heat transfer coefficient. wall The temperature of the wind turbine wall. for Air temperature at time A ex Let T be the heat exchange area between the wind turbine and the air. Then, the total energy over time period T can be expressed as the energy per unit time over time T. Using the law of conservation of energy, we can obtain the energy conversion relationships: ;

[0092] The energy matrix W obtained over time T is expressed as: ;

[0093] Considering the losses during energy conversion, a compensation coefficient matrix is ​​set. To approximate the energy conservation equation, where C w To compensate for the loss of wind energy converted into mechanical energy, C is the compensation coefficient. r C is the compensation coefficient for capacitance conversion. qin For and C qex Let be the energy conversion compensation coefficients for heat absorption and heat release, respectively. Assuming that the compensation coefficients remain constant within the time period T, when there is sufficient data, a system of equations can be constructed using Newton's Cotes formula, and then the corresponding compensation coefficients can be solved using the method of undetermined coefficients. ;

[0094] in This is the inverse matrix of W. To make the physical model more closely related to the neural network model, the original input defined in step 3 is... The measurement value at the last time step is obtained by inputting it into the RNN model. And with the target power The specific method for splicing is as follows: ;

[0095] And recalculated ,in The energy matrix is ​​obtained by calculating using the spliced ​​data, and the final predicted electrical energy is:

[0096]

[0097] in The electrical energy is calculated using auxiliary variables and weighting functions in the physical model.

[0098] However, due to the difficulty in fitting the variables to the target electrical energy in wind power data, the weight function obtained by simply inverting the undetermined coefficients method cannot reconstruct the accurate value well. In order to construct a power generation prediction model based on the mechanism model of wind power generation, this invention uses a KAN network to reconstruct the physical weight function. This theorem states that any continuous multivariate function containing n inputs can be decomposed into a linear combination of finite univariate functions. Therefore, this invention uses KAN approximation to obtain the mechanism model of wind power generation: ;

[0099] in This represents the mechanism function of wind power generation fitted by the weighted KAN network model. and This represents a trainable inner unary function. For trainable outer combination functions, Since represents the p-th feature variable, q represents the q-th univariate function. Because KAN decomposes complex multivariate functions into linear combinations of several univariate functions, it avoids the coupling of complex variables. Inner function and The outer function is able to focus on capturing the physical relationship between each input feature variable and the output variable. Its role is to integrate the contributions of multiple features, providing a basis for obtaining the expected physical output.

[0100] Wind power generation exhibits strong temporal dynamics. While the above mechanism data fusion modeling process extracts the nonlinear relationship between input and output, it does not consider these temporal dynamics. Given that energy is the accumulation of power over time, therefore, in terms of time... The energy within the time frame can be obtained by integrating power over time. Therefore, we propose an average optimization method based on numerical integration, which first integrates the power over time. The power data within the time frame is numerically integrated to fuse the power time-time dynamic characteristics from multiple time steps into energy characteristics. Since the data consists of discrete values ​​sampled at fixed time intervals, this paper uses the following Newton-Cotes formula to solve the numerical integral, thereby approximating the integral term in the above formula. The Newton-Cotes formula is shown below. ;

[0101] Where a and b are the endpoints of the region to be multiplied, the region to be multiplied will be divided into H equal parts. for For any feature variable in the equation, the k-th division point has numerical value. , These are the corresponding coefficients of Newton-Cotes' formula. Let be the function that needs to be multiplied.

[0102] The electrical energy E can then be calculated using the following formula: ;

[0103] Where E represents the accumulated electrical energy within the time period T. express The active power of a time step, where H represents the number of time steps within the time period. These are the constant coefficients of the corresponding Newton-Cotes formula.

[0104] To obtain the predicted value of accumulated electrical energy Design a KAN feature processing model to process the input sequence The physical feature matrix is ​​obtained after processing by the inner function of KAN. Then, the same numerical integration method is applied to each variable of the input sequence to obtain the integral feature matrix. This is used to characterize the time-cumulative properties of historical data. The physical feature matrix and the integral feature matrix are then concatenated and input into the outer function to obtain the predicted value of the accumulated electrical energy. .

[0105] Finally, to ensure physical consistency, the energy residual loss is defined as follows: ;

[0106] Where N represents the number of historical data intervals. and Let represent the predicted and actual electrical energy values ​​at time k, respectively.

[0107] (5) Source domain pre-training and physics-guided transfer: The pre-trained model trained on the source domain is saved, the physical model is frozen and then transferred to the target domain. The pseudo-labels generated on the target domain are optimized using the adversarial training mode, and the model parameters are fine-tuned. Finally, the model is tested on the test set of the target domain.

[0108] Step 5.1: Source Domain Model Training:

[0109] During the source domain model training phase, to meet the requirements of temporal dynamic modeling, an RNN model was used to extract the spatiotemporal features of wind power data. For the input historical auxiliary variable sequence... At each time step, the RNN updates the hidden state h using the following recursive relationship. k : ;

[0110] in, For input, W xh W hh These are the input and state transition weight matrices, respectively, b h For bias terms, This represents the tanh activation function. The final model output is... Through the following calculations: ;

[0111] Among them W hy To output the weight matrix, b y For output bias terms, This represents the output mapping function of the fully connected layer.

[0112] The RNN model described above implements the mapping relationship from the input sequence to the output sequence. To further train the model and make the prediction results closer to the true values, the following loss function is introduced to constrain the prediction error: ;

[0113] Then, the predicted output of the RNN model is... with input history auxiliary variable sequence After splicing, input the KAN mechanism model defined in step 4 and obtain the energy residual R. S Finally, the loss function for training the source domain model is: ;

[0114] in These are the hyperparameters used in the balancing model.

[0115] Step 5.2: Freeze the migration and fine-tune the target domain model:

[0116] Due to significant differences in geographical environment, meteorological conditions, and operational characteristics among different wind farms, directly applying the source domain model to the target domain can lead to performance degradation in prediction. To reduce cross-domain distribution differences, this study constructs a target domain model based on the frozen physical migration method. Historical data for the target domain... First, the RNN module is based on historical auxiliary variables. Predicted value of output active power In the source domain task, the RNN module aims to capture the nonlinear mapping relationship between variables and does not predict the power generation at the next time step. Therefore, during the transfer process, to retain the general features learned in the source domain model, this study uses parameter transfer to freeze most of the network parameters and only retrains some transferable layers on the target domain data to achieve domain adaptation of the model. Fine-tuning is performed on the fully connected layers in the transferred RNN module, and the loss function is constructed using the target domain data. ;

[0117] in , These are the predicted value and the actual power output by the RNN module at time step j in the target domain, respectively. Then, the predicted output of the RNN model... with input history auxiliary variable sequence After splicing, the data is input into the KAN mechanism model, and the energy residual R is obtained. T Finally, the loss function for training the source domain model is: ;

[0118] in These are the hyperparameters used to balance the model. Then, the RNN module and the KAN mechanism module are fixed, and the prediction output of the RNN model is... with input history auxiliary variable sequence After concatenation, the data is input into the KAN mechanism model to obtain pseudo-labels. However, this pseudo-label cannot accurately adapt to the mapping relationship between features. Therefore, to enhance domain adaptation, an adversarial mechanism is constructed and optimized using the following loss function: ;

[0119] in It is the predicted power after training in the source domain and transferring it to the target domain. These are pseudo-labels generated by the physical model on the target domain. This represents the discriminator model.

[0120] For the KAN mechanism module, the following loss function is used to update it so that it is difficult to distinguish from the source domain from the perspective of the discriminator, and finally the domain-invariant representation is achieved.

[0121] ;

[0122] The optimization goal of overall adversarial training is to ;

[0123] in To balance the hyperparameters of adversarial training, the KAN mechanism module and discriminator are alternately updated to achieve real-time correction of pseudo-labels in the target domain, thereby enabling accurate adversarial transfer. Specifically, the designed neural network module includes three RNN layers, three activation layers, and one fully connected layer. The trainable physical model module includes a compensation coefficient fitting KAN network and a physical feature processing network. During training in the source domain, the neural network model, compensation coefficient calculation network, and physical feature processing network are trained simultaneously, and their parameters are saved. When transferring to the target domain, both the saved neural network model and physical model are transferred, fine-tuned in the target domain, and the trainable parameters in the RNN layers, compensation coefficient calculation network, and physical feature processing network are frozen. Only the fully connected layer in the neural network model is trained. This helps improve the model's stability and transfer robustness in the new domain, especially when the target domain lacks sufficient physical labels or is difficult to model physically.

[0124] (6) Model performance evaluation: Evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The model is evaluated and compared with other models.

[0125] Use MAE, RMSE, and R 2 Metrics are used to evaluate the performance of the model.

[0126] ;

[0127] in This represents the mean of all true values. Indicates actual power. The predicted power is represented by . To demonstrate the robustness of the model, the model was tested on two different datasets as the source domain. The results of each model trained on the training set and tested on the test set are shown in Table 1. Specifically, on the experimental dataset, after the RNN model was improved using the method of this invention, the RMSE and MAE values ​​decreased by 10.61% and 44.25%, respectively. 2 The performance was improved by 9.81%, and a comparison was made with an Unsupervised Domain Adaptation (UDA) model based on adversarial training. The main function of the UDA model is to make the features of the source and target domains as close as possible to each other in the latent space through feature alignment, and to use a shared representation learning feature extractor to extract common information between the two domains, avoiding overfitting to the source domain. This enables unsupervised transfer learning through source domain knowledge and adversarial training, even if the target domain has no label. The results show that the method of this invention can effectively improve the performance of the prediction model. Figures 2-4 As shown in the scatter plot and line graph, after using the method of the present invention, the predicted value of wind power generation is closer to the actual value and has a lower absolute prediction error.

[0128] Table 1 Model Performance Comparison ;

[0129] This invention proposes a wind power prediction method that integrates physical modeling, feature selection, and transfer mechanisms. By introducing the energy conservation equation as a physical constraint, the physical consistency and accuracy of the prediction results are improved. The maximum information coefficient is used to screen key features, and a KAN network is combined to extract physically meaningful nonlinear structural features, enhancing the model's robustness and generalization ability. Furthermore, to address the difficulty in collecting target domain labels, a transfer strategy that freezes the physical model is adopted, achieving high prediction performance and stability even under label-less conditions.

[0130] This invention combines maximum information coefficient feature selection, physical knowledge modeling, and feature processing layer freeze migration mechanism to create an unsupervised short-term future power prediction method, which is suitable for wind power prediction scenarios where power is difficult to measure.

[0131] 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 few-sample wind power prediction method integrating mechanism transfer modeling, characterized in that, Includes the following steps: Step 1) Obtain the original sample data and filter out the auxiliary variables that are related to the predictor variables; Step 2) Preprocessing of raw sample data and partitioning of the dataset: The original sample data is time-series processed by using a moving window, and then the time-series sample data is divided into training set and test set; and the training set data and test set data are normalized. Step 3) Domain definition and prediction objectives: Based on the time-series data, determine the time-series data range of the source and target domains, and clarify the prediction tasks for both. Step 4) KAN mechanism fusion model: A simplified energy conservation equation is established. The measured power is concatenated with the target power using a recurrent neural network (RNN) model. The concatenated data is then used to calculate various energy parameters in the wind power generation process, and the calculation results are corrected using a compensation coefficient. A KAN network based on the Kolmogorov representation theorem is introduced. The selected auxiliary variables are nonlinearly projected dimension by dimension to obtain new auxiliary features. The residual loss is calculated using the constructed simplified energy conservation equation, and the model parameters are optimized on the training set. 5) Source domain pre-training and physics-guided transfer: The pre-trained model trained on the source domain is saved, the physical model is frozen and then transferred to the target domain. The pseudo-labels generated on the target domain are optimized using adversarial training mode, and the model parameters are fine-tuned. Finally, the model is tested on the test set of the target domain.

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 3) is as follows: The source domain dataset is Include Historical data at each time step, including Let F be the auxiliary variables at time k. This represents the power generation at time k; the target task of the source domain is to utilize... Auxiliary variables at each time step Measure the power generation at the last time step For those containing time step The target domain data, the task is to utilize historical auxiliary variables. predict Power generation at any time .

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, The specific process of step 4) is as follows: Step 4.1: Establish a simplified energy conservation equation: Based on the energy conversion principle of wind turbines, combined with the calculation formulas for air kinetic energy, absorbed heat energy, and convective heat transfer, a simplified energy conservation equation is obtained, thus yielding the energy matrix W over time T; the original input... The measurement value at the last time step is obtained by inputting it into the RNN model. , with target power The energy matrix is ​​recalculated by splicing the components and introducing a compensation coefficient matrix C. ; Step 4.2: Construct a power generation prediction model based on the mechanism model of wind power generation. Use a KAN network to reconstruct the physical weight function to obtain the mechanism model of wind power generation. ； in This represents the mechanism function of wind power generation fitted by the weighted KAN network model. and This represents a trainable inner unary function. For trainable outer combination functions, Since q represents the p-th characteristic variable, q represents the q-th unary function; Step 4.3: Construct an average optimization model based on numerical integration. First, consider the time... The power data within the time step is numerically integrated to fuse the power time-time dynamic characteristics of multiple time steps into energy characteristics. The numerical integral is then solved using Newton-Cotes' formula, resulting in the following formula for calculating electrical energy E: ； Where E represents the accumulated electrical energy within the time period T. express The active power of a time step, where H represents the number of time steps within the time period. These are the constant coefficients of the corresponding Newton-Cotes formula; Step 4.4: Construct the KAN feature processing model to obtain the predicted value of accumulated electrical energy. : Input sequence The physical feature matrix is ​​obtained after processing by the inner function of KAN. Then, the same numerical integration method is applied to each variable of the input sequence to obtain the integral feature matrix. ; The physical feature matrix and the integral feature matrix are then concatenated and input into the outer function to obtain the predicted value of the accumulated electrical energy. ; Define energy residual loss: ； Where N represents the number of historical data intervals. and Let represent the predicted and actual electrical energy values ​​at time k, respectively.

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, The specific process of step 5) is as follows: Step 5.1: Source Domain Model Training: During the source domain model training phase, an RNN model is used to extract the spatiotemporal features of wind power data; for the input historical auxiliary variable sequence... At each time step, the RNN updates the hidden state using a recursive relationship. The final model output is ; The following loss function is introduced to constrain the prediction error: ； Then, the predicted output of the RNN model is... with input history auxiliary variable sequence After splicing, input the KAN mechanism fusion model from step 4 to obtain the energy residual R. S Finally, the loss function for training the source domain model is: ； in, These are the hyperparameters used in the equilibrium model; Step 5.2: Freeze the migration and fine-tune the target domain model: To reduce cross-domain distribution differences, a target domain model is constructed based on the frozen physical migration method: for historical data of the target domain First, the RNN module is based on historical auxiliary variables. Predicted value of output active power In the source domain task, the RNN module aims to capture the nonlinear mapping relationship between variables, but does not predict the power generation at the next time step; During the transfer process, parameter transfer is used to retrain only some transferable layers on the target domain data; the fully connected layers in the transferred RNN module are fine-tuned, and the loss function is constructed using the target domain data. ； in, , These are the predicted value and the actual power output by the RNN module at time step j in the target domain, respectively. Then, the predicted output of the RNN model is... with input history auxiliary variable sequence After splicing, the data is input into the KAN mechanism fusion model, and the energy residual R is obtained. T Finally, the loss function for training the source domain model is: ； in, These are the hyperparameters used in the equilibrium model; Then, fixing the RNN module and the KAN mechanism module, the prediction output of the RNN model is... with input history auxiliary variable sequence After concatenation, the data is input into the KAN mechanism fusion model to obtain pseudo-labels. ; To enhance domain adaptation, an adversarial mechanism is constructed and optimized using the following loss function: ； in, It is the predicted power after training in the source domain and transferring it to the target domain. These are pseudo-labels generated by the physical model on the target domain. Represents the discriminator model; For the KAN mechanism fusion model, the following loss function is used for updating, ultimately achieving a domain-invariant representation; ； The optimization goal of the overall adversarial training is: ； in, To balance the hyperparameters of adversarial training; By alternately updating the KAN mechanism fusion model and the discriminator model, real-time correction of pseudo-labels can be achieved in the target domain, thereby enabling accurate adversarial transfer.

Images