Wind turbine wake field modeling and prediction method and device

Abstract

This application provides a method and apparatus for modeling and predicting the wake field of a wind turbine, addressing the difficulty in balancing accuracy, efficiency, and physical consistency in existing wake prediction methods. The method includes: constructing a deep neural network model, outputting predicted values ​​of corresponding flow field physical quantities; calculating the residuals of the physical equations based on the hydrodynamic control equations followed by the wind turbine wake, using the output of the deep neural network and automatic differentiation techniques, and constructing a composite loss function including data fitting loss terms, physical constraint loss terms, boundary condition loss terms, and initial condition loss terms; training the deep neural network model using a training dataset by minimizing the composite loss function to obtain a wind turbine wake field prediction model; and using the wind turbine wake field prediction model to simulate and predict the wake field of a target wind turbine.

Description

Technical Field

[0001] This application relates to the field of wind power generation, specifically to a method and apparatus for modeling and predicting the wake field of a wind turbine based on a physical information neural network. Background Technology

[0002] Wind energy, as a clean and renewable energy source, occupies an important strategic position in the global energy transition. During wind farm operation, the wake effect generated by the rotation of upstream wind turbines significantly reduces the effective wind speed of downstream turbines, thereby affecting the overall power generation efficiency and exacerbating turbine fatigue loads. Therefore, achieving accurate and efficient wake field reconstruction and prediction is a key technical challenge for optimizing wind farm layout, increasing power generation, and ensuring operational safety.

[0003] Currently, the mainstream wake restoration methods mainly suffer from the following technical limitations:

[0004] Experimental measurement method: This method involves conducting on-site observations by deploying wind measurement towers, lidar, and other equipment at the wind farm. While the data obtained using this method is accurate and reliable, it suffers from inherent drawbacks such as high cost, long deployment time, and limited spatial coverage, making it difficult to meet the needs of large-scale panoramic monitoring and real-time analysis of wind farms.

[0005] Computational fluid dynamics simulation: This method performs high-fidelity simulations of the wake field based on numerical solutions to the Navier-Stokes equations. While it achieves high accuracy under specific conditions, it relies on fine mesh generation and iterative solutions, resulting in significant computational resource consumption and lengthy processing times. Therefore, it is unsuitable for real-time state awareness and rapid optimization decision-making scenarios in wind farms.

[0006] Theoretical / semi-empirical model methods, such as the Jensen and Larsen models, establish analytical expressions based on simplified fluid dynamics assumptions. These methods are computationally fast, but the high degree of model simplification leads to significant prediction errors under real-world conditions such as complex terrain, atmospheric boundary layer variations, and wind turbine yaw, resulting in insufficient generalization and adaptability.

[0007] In recent years, with the development of big data and artificial intelligence technologies, purely data-driven methods based on wind farm monitoring and data acquisition systems have emerged, such as using deep neural networks to directly learn wake patterns from historical data. While these methods improve computational efficiency, they inherently lack physical constraints, exhibiting "black box" characteristics. Prediction reliability drops sharply under operating conditions outside the training data distribution, and it is difficult to guarantee that the prediction results conform to fundamental physical laws such as mass and momentum conservation, resulting in limited engineering credibility and robustness.

[0008] Therefore, the industry urgently needs a new technical solution that can simultaneously meet the following requirements:

[0009] (1) It can make full use of the unique multi-source heterogeneous data in the wind power field (such as SCADA operation data, wind turbine design parameters, and meteorological monitoring data) for efficient fusion and feature extraction;

[0010] (2) It can effectively integrate the basic principles of fluid mechanics during the learning process to ensure that the prediction results conform to the basic physical laws;

[0011] (3) It has computational efficiency close to that of theoretical models and prediction accuracy close to that of CFD methods, which can meet the engineering requirements of real-time monitoring and rapid optimization of wind farms.

[0012] (4) It can handle wake prediction problems under complex working conditions and has good generalization ability and engineering applicability.

[0013] However, no complete solution in the current technology can simultaneously meet all the above requirements. In particular, existing technologies have significant shortcomings in how to effectively integrate fluid dynamics equations with deep learning frameworks and how to construct a dedicated model architecture for wind turbine wake prediction. Summary of the Invention

[0014] This application provides a method and apparatus for modeling and predicting the wake field of a wind turbine generator, in order to solve the problem that existing wake prediction methods are difficult to balance between accuracy, efficiency and physical consistency.

[0015] The technical solution of this invention is as follows:

[0016] This application provides a method for modeling and predicting the wake field of wind turbines based on physical information neural networks, including:

[0017] Multi-source observation data of the flow field where the wind turbine is located and wind turbine design parameters are obtained. An input feature vector containing spatiotemporal coordinates, incoming flow conditions and wind turbine characteristic parameters is constructed. The real flow field physical quantity data corresponding to the input feature vector are obtained and together they form a training dataset.

[0018] A deep neural network model is constructed, with the input feature vector as its input and the corresponding predicted values ​​of flow field physical quantities as its output. Based on the fluid dynamics control equations followed by the wind turbine wake, the residuals of the physical equations are calculated using the output of the deep neural network and automatic differentiation techniques. A composite loss function is constructed, comprising a data fitting loss term, a physical constraint loss term, a boundary condition loss term, and an initial condition loss term. The data fitting loss term is calculated based on the difference between the actual flow field physical quantity data and the predicted values ​​of the flow field physical quantities.

[0019] Using the training dataset, the deep neural network model is trained by minimizing the composite loss function to obtain a wind turbine wake field prediction model;

[0020] The wind turbine wake field prediction model is used to simulate and predict the wake field of the target wind turbine.

[0021] Preferably, the input feature vector includes at least spatial coordinates x, y, z, time coordinate t, incoming wind speed U∞, wind direction θ, wind turbine diameter D, thrust coefficient Ct, and power coefficient Cp.

[0022] Preferably, the deep neural network model includes two fully connected hidden layers connected in sequence, each containing 100 neurons, and the activation function is a hyperbolic tangent function or a Swish function; the output layer of the deep neural network model directly outputs the predicted values ​​of the three-dimensional velocity components u, v, w and pressure p.

[0023] Preferably, the physical equation residuals include the continuity equation residuals R. cont With the residual R of the momentum equation x ,R y ,R z It automatically differentiates the partial derivatives of the predicted flow field physical quantities output by the deep neural network model and substitutes them into the following Navier-Stokes equations to obtain:

[0024]

[0025]

[0026] Where u, v, w are the predicted values ​​of the fluid velocity components in the x, y, z directions output by the deep neural network model, p is the predicted value of the fluid pressure output by the deep neural network model, ρ is the fluid density, ν is the fluid kinematic viscosity, and fx, fy, fz are the volume force components used to simulate the force of the wind turbine blades on the airflow.

[0027] Preferably, the initial condition loss term The expression is:

[0028] ;

[0029] in, The total number of initial condition points, where each initial condition point is a spatiotemporal coordinate point with a known true flow field state at the initial time t=0; m is the index of the initial condition point. , and These are the velocity components of the fluid in the x, y, and z directions predicted by the neural network model at the m-th initial condition point when the initial time t=0; and These represent the true velocity components in the x, y, and z directions of the fluid at the m-th initial condition point recorded in the training dataset at the initial time t=0.

[0030] Preferably, the physical constraint loss term The expression is:

[0031] ;

[0032] Let i be the number of randomly sampled configuration points within the spatiotemporal computational domain of the wind turbine wake field, and let i be the index of the configuration point. For the continuity equation residual at the i-th location point, , and Let be the momentum equation residuals in the x, y, and z directions at the i-th placement point, respectively. for The weighting coefficients, , and They are respectively , and The weighting coefficients.

[0033] Preferably, the data fitting loss term The expression is:

[0034] ;

[0035] in, The total number of data points in the training dataset. , and These represent the velocity components of the fluid in the x, y, and z directions predicted by the neural network model at the j-th data point. This represents the fluid pressure predicted by the neural network model at j data points. , and Let x, y, and z represent the actual velocity components in the x, y, and z directions corresponding to the j-th data point, respectively. Let represent the actual pressure corresponding to the j-th data point.

[0036] Preferably, the boundary condition loss term The expression is:

[0037]

[0038] in, The total number of boundary condition points, where each boundary condition point is a spatiotemporal coordinate point with a known flow field state on the boundary of the computational domain of the wind turbine wake field; k is the index of the boundary condition point. , and These are the velocity components of the fluid in the x, y, and z directions predicted by the neural network model at the k-th boundary condition point. , and These are the x, y, and z-direction boundary velocity components of the fluid at the k-th boundary condition point, as preset according to the physical laws of the flow field.

[0039] Preferably, the step of training the deep neural network model by minimizing the loss function using the training dataset includes:

[0040] The weights and bias parameters of the deep neural network model are randomly initialized and assigned values;

[0041] Training batches are sampled from the training dataset, and multiple rounds of iterative optimization are performed. Each iteration includes: inputting the spatiotemporal coordinates of the training batches into the model for forward propagation to obtain predicted values ​​of flow field physical quantities; calculating the partial derivatives of the predicted values ​​using automatic differentiation; calculating the composite loss function based on the predicted values ​​and their partial derivatives, combined with the real physical quantity labels in the training dataset; calculating the gradient of the loss function with respect to the model parameters through backpropagation, and updating the model parameters using an adaptive moment estimation optimization algorithm or an L-BFGS optimization algorithm.

[0042] During or after the training iteration optimization process, the model performance is evaluated using the validation dataset; hyperparameters are adjusted based on the validation results, including at least one of network depth, network width, activation function type, weight coefficients of each item in the composite loss function, learning rate, and batch size; the training iteration optimization steps are continued or re-executed based on the adjusted hyperparameters.

[0043] When the preset convergence condition is met, training stops and the current model parameters are output to obtain the wind turbine wake field prediction model; the convergence condition is: the value of the composite loss function is lower than the preset threshold, or the rate of change of the composite loss function in multiple consecutive iterations is less than the stability threshold, or the preset maximum number of iterations is reached.

[0044] This invention also provides a wind turbine wake field modeling and prediction device based on physical information neural networks, comprising:

[0045] The acquisition module is used to acquire multi-source observation data of the flow field where the wind turbine is located and the design parameters of the wind turbine, construct an input feature vector containing spatiotemporal coordinates, incoming flow conditions and wind turbine characteristic parameters, and acquire the real flow field physical quantity data corresponding to the input feature vector, which together constitute the training dataset.

[0046] The model building module is used to construct a deep neural network model, whose input is the input feature vector and whose output is the corresponding predicted value of the flow field physical quantity. Based on the fluid dynamics control equations followed by the wind turbine wake, the residuals of the physical equations are calculated using the output of the deep neural network and automatic differentiation techniques. A composite loss function is constructed, including a data fitting loss term, a physical constraint loss term, a boundary condition loss term, and an initial condition loss term. The data fitting loss term is calculated based on the difference between the actual flow field physical quantity data and the predicted value of the flow field physical quantity.

[0047] The model training module is used to train the deep neural network model by minimizing the composite loss function using the training dataset to obtain a wind turbine wake field prediction model.

[0048] The application module is used to use the wind turbine wake field prediction model to simulate and predict the wake field of the target wind turbine.

[0049] The beneficial effects of this invention are as follows:

[0050] By using the Navier-Stokes equations as soft constraints and embedding them into the loss function of the neural network through automatic differentiation, the model's predictions naturally satisfy fundamental physical laws such as mass and momentum conservation, fundamentally guaranteeing physical consistency and overcoming the shortcomings of pure data-driven methods, such as black-box prediction and poor physical interpretability. Simultaneously, the forward propagation inference speed of deep neural networks is extremely fast; once trained, the prediction time for a new operating condition's wake field can be several orders of magnitude faster than traditional CFD simulations, meeting the efficiency requirements for real-time monitoring and rapid optimization of wind farms. Therefore, this method achieves a unity of accuracy close to CFD simulation, speed comparable to theoretical models, and physical consistency far exceeding pure data-driven methods.

[0051] Furthermore, traditional pure data-driven models heavily rely on the distribution of training data, resulting in significant performance degradation under conditions such as scarce or unseen data, extreme wind conditions, complex terrain, and wind turbine yaw. In this solution, the physical constraint loss term acts as a physical regularization. Even when data is scarce or unavailable for specific operating conditions, the constraints of physical laws guide the neural network to extrapolate and predict in a physically reasonable direction, greatly enhancing the model's generalization ability and robustness to extreme conditions. In addition, the input features incorporate wind turbine design parameters (D, Ct, Cp) and real-time inflow conditions, enabling the model to adapt to different wind turbine models and real-time wind conditions, further improving its engineering applicability.

[0052] The study explicitly proposes integrating SCADA measured data, anemometer data, and CFD simulation data to construct a training set. SCADA and anemometer data provide key information under real-world operating conditions, ensuring the model's close alignment with reality; CFD simulation data provides spatially continuous and physically complete high-fidelity flow field information, compensating for the spatial sparsity and incomplete physical quantities of measured data. By constructing an input vector containing wind turbine characteristic parameters, the model can learn the profound impact of the wind turbine's own aerodynamic characteristics on the wake. Attached Figure Description

[0053] Figure 1 A flowchart of the wind turbine wake field modeling and prediction method in the embodiments of this application;

[0054] Figure 2 This is a flowchart illustrating the wind turbine wake field modeling and prediction method in the embodiments of this application. Detailed Implementation

[0055] Reference Figure 1 This application provides a method for modeling and predicting the wake field of a wind turbine based on a physical information neural network, including:

[0056] S101, acquire multi-source observation data of the flow field where the wind turbine is located and wind turbine design parameters, construct an input feature vector containing spatiotemporal coordinates, incoming flow conditions and wind turbine characteristic parameters, and acquire the real flow field physical quantity data corresponding to the input feature vector, together forming a training dataset;

[0057] S102, Construct a deep neural network model, whose input is the input feature vector and whose output is the corresponding predicted value of the flow field physical quantity; Based on the fluid dynamics control equations followed by the wind turbine wake, calculate the physical equation residuals based on the output of the deep neural network and automatic differentiation technology, and construct a composite loss function including a data fitting loss term, a physical constraint loss term, a boundary condition loss term, and an initial condition loss term; wherein the data fitting loss term is calculated based on the difference between the actual flow field physical quantity data and the predicted value of the flow field physical quantity;

[0058] S103, Using the training dataset, the deep neural network model is trained by minimizing the composite loss function to obtain a wind turbine wake field prediction model;

[0059] S104, The wind turbine wake field prediction model is used to simulate and predict the wake field of the target wind turbine.

[0060] In this embodiment of the application, in step S101, multi-source observation data collection is first performed, specifically including:

[0061] Data Collection: Historical and real-time SCADA data from wind farms, as well as meteorological tower data, are collected. Real-time SCADA data includes, but is not limited to, wind speed, wind direction, turbine power, and rotational speed. Meteorological tower data includes, for example, data from a meteorological tower located within an unobstructed sector (e.g., within ±22.5°). The raw data undergoes cleaning and interpolation. Cleaning aims to remove outliers and missing values ​​caused by sensor malfunctions or communication interruptions. Interpolation is performed by using data from different heights on the same tower or mesoscale meteorological data for missing data.

[0062] Wind turbine parameter collection includes: rotor diameter (D), rated power, thrust coefficient (Ct), and power coefficient (Cp). Among them, the thrust coefficient (Ct) and power coefficient (Cp) can be calculated by combining measured SCADA data (wind speed U, power P, air density ρ) with wind turbine aerodynamic theory.

[0063] Simulation data collection: Based on actual wind conditions and fan parameters, high-fidelity three-dimensional velocity field (u, v, w) and pressure field (p) are generated through computational fluid dynamics (CFD) simulation.

[0064] Based on the effective data processed above, an input feature vector is constructed; wherein the input feature vector includes at least the spatial coordinates x, y, z, the time coordinate t, the incoming wind speed U∞, the wind direction θ, the wind turbine diameter D, the thrust coefficient Ct, and the power coefficient Cp.

[0065] Furthermore, for each input feature vector, the corresponding real flow field physical quantity data is matched as the label for supervised learning; among them, for SCADA / wind measurement tower data points, the real values ​​are mainly the measured wind speed components (or derived values); for CFD simulation data points, the real values ​​are the complete velocity field (u, v, w) and pressure field (p) output by the simulation.

[0066] The above input feature vectors are paired with the corresponding real flow field physical quantity data to form a large number of samples, which together constitute the training dataset for model training.

[0067] In this embodiment, the constructed deep neural network model includes two fully connected hidden layers connected in sequence, each containing 100 neurons, and the activation function is either the hyperbolic tangent function or the Swish function; the output layer of the deep neural network model directly outputs the predicted values ​​of the three-dimensional velocity components u, v, w and pressure p.

[0068] In step S102, the deep neural network model uses a feedforward neural network (FNN) or residual network (ResNet) as its basic architecture, takes the above-mentioned input feature vector as input, and outputs the predicted values ​​of the flow field physical quantities.

[0069] In this embodiment of the application, the physical equation residual includes the continuity equation residual R. cont With the residual R of the momentum equation x ,R y ,R z The continuity equation residual R cont With the residual R of the momentum equation x ,R y ,R z The partial derivatives of the predicted flow field physical quantities output by the deep neural network model are calculated by automatic differentiation and then substituted into the Navier-Stokes equations.

[0070] Among them, the residual R of the continuity equation cont :

[0071]

[0072] For the residual of the momentum equation in the x-direction:

[0073]

[0074] Similarly, the y and z directions can be obtained:

[0075]

[0076] All first and second-order partial derivatives are automatically calculated using the self-differentiation technique of a neural network, ensuring both accuracy and computational efficiency. From this, the complete momentum equation residual R can be obtained. x ,R y ,R z The expression is:

[0077]

[0078] Where u, v, and w are the predicted velocity components of the fluid in the x, y, and z directions output by the deep neural network model, respectively; p is the predicted fluid pressure output by the deep neural network model; ρ is the fluid density; ν is the fluid kinematic viscosity; fx, fy, and fz are the volume force components used to simulate the force exerted by the wind turbine blades on the airflow. In wind turbine wake simulation, volume force is usually used to represent the force exerted by the wind turbine blades on the airflow (i.e., thrust). For example, in the wind turbine wake model, we usually apply a drag (thrust) at the wind turbine location to simulate the wind turbine extracting momentum. This force usually acts in the x direction (mainstream direction), so fx may not be zero, while fy and fz are usually zero.

[0079] , , These represent the rates of change of velocity over time in the x, y, and z directions, respectively, reflecting the unsteady characteristics of the flow, lateral velocity fluctuations, and vertical velocity changes.

[0080] In this embodiment of the application, the initial condition loss term The expression is:

[0081] ;

[0082] in, The total number of initial condition points, where each initial condition point is a spatiotemporal coordinate point with a known true flow field state at the initial time t=0; m is the index of the initial condition point. , and These are the velocity components of the fluid in the x, y, and z directions predicted by the neural network model at the m-th initial condition point when the initial time t=0; and These represent the true velocity components in the x, y, and z directions of the fluid at the m-th initial condition point recorded in the training dataset at the initial time t=0.

[0083] In this embodiment of the application, the physical constraint loss term The expression is:

[0084] ;

[0085] Let i be the number of randomly sampled configuration points within the spatiotemporal computational domain of the wind turbine wake field, and let i be the index of the configuration point. For the continuity equation residual at the i-th location point, , and Let be the momentum equation residuals in the x, y, and z directions at the i-th placement point, respectively. for The weighting coefficients, , and They are respectively , and The weighting coefficients.

[0086] In this embodiment of the application, the data fitting loss term The expression is:

[0087] ;

[0088] in, The total number of data points in the training dataset. , and These represent the velocity components of the fluid in the x, y, and z directions predicted by the neural network model at the j-th data point. This represents the fluid pressure predicted by the neural network model at j data points. , and Let x, y, and z represent the actual velocity components in the x, y, and z directions corresponding to the j-th data point, respectively. Let represent the actual pressure corresponding to the j-th data point.

[0089] In this embodiment of the application, the boundary condition loss term The expression is:

[0090]

[0091] in, The total number of boundary condition points, where each boundary condition point is a spatiotemporal coordinate point with a known flow field state on the boundary of the computational domain of the wind turbine wake field; k is the index of the boundary condition point. , and These are the velocity components of the fluid in the x, y, and z directions predicted by the neural network model at the k-th boundary condition point. , and These are the x, y, and z-direction boundary velocity components of the fluid at the k-th boundary condition point, as preset according to the physical laws of the flow field.

[0092] Boundary conditions include, but are not limited to, entrance boundary, exit boundary, wall boundary, and far-field boundary.

[0093] In this embodiment of the application, step S103, which involves training the deep neural network model by minimizing the loss function using the training dataset, includes:

[0094] The weights and bias parameters of the deep neural network model are randomly initialized and assigned values;

[0095] Training batches are sampled from the training dataset, and multiple rounds of iterative optimization are performed. Each iteration includes: inputting the spatiotemporal coordinates of the training batches into the model for forward propagation to obtain predicted values ​​of flow field physical quantities; calculating the partial derivatives of the predicted values ​​using automatic differentiation; calculating the composite loss function based on the predicted values ​​and their partial derivatives, combined with the real physical quantity labels in the training dataset; calculating the gradient of the loss function with respect to the model parameters through backpropagation, and updating the model parameters using an adaptive moment estimation optimization algorithm or an L-BFGS optimization algorithm.

[0096] During or after the training iteration optimization process, the model performance is evaluated using the validation dataset; hyperparameters are adjusted based on the validation results, including at least one of network depth, network width, activation function type, weight coefficients of each item in the composite loss function, learning rate, and batch size; the training iteration optimization steps are continued or re-executed based on the adjusted hyperparameters.

[0097] When the preset convergence condition is met, training stops and the current model parameters are output to obtain the wind turbine wake field prediction model; the convergence condition is: the value of the composite loss function is lower than the preset threshold, or the rate of change of the composite loss function in multiple consecutive iterations is less than the stability threshold, or the preset maximum number of iterations is reached.

[0098] In step S104, the wind turbine wake field prediction model, which has been trained, converged, and validated, is deployed to the wind farm monitoring system. By inputting real-time SCADA data and wind turbine state parameters, the three-dimensional velocity and pressure fields of the entire wind farm can be quickly predicted, achieving the following:

[0099] Real-time wake assessment: Near real-time visualization of the degree of wake impact on downstream wind turbines.

[0100] Increased power generation: Provides data support for optimized wind farm scheduling (such as yaw control and torque control) to reduce wake loss and increase overall power generation.

[0101] Operation monitoring and early warning: Identify abnormal flow field conditions and provide early warning for the safe operation of the wind turbine.

[0102] The method described in this application embeds the Navier-Stokes equations as hard constraints into the learning process, enabling the model prediction results to not only fit the training data but also strictly follow the basic laws of fluid mechanics. Even in sparse data regions, it can ensure physical rationality, and the prediction accuracy and generalization ability are far superior to those of pure data-driven models.

[0103] The wind turbine wake field prediction model, once trained, requires only one forward propagation when performing flow field inference. Its calculation speed is several orders of magnitude faster than traditional CFD simulation, which can meet the needs of real-time or near-real-time simulation of wind farms.

[0104] A complete technical framework integrating SCADA measured data and CFD simulation data has been constructed, resolving the contradiction between limited data and high precision requirements in industrial scenarios. The model can be directly integrated into existing wind farm SCADA systems, providing a powerful technical tool for power generation assessment, optimized scheduling, and safety monitoring.

[0105] Reference Figure 2 This application also provides a wind turbine wake field modeling and prediction device based on physical information neural network, including:

[0106] The acquisition module 101 is used to acquire multi-source observation data of the flow field where the wind turbine is located and the wind turbine design parameters, construct an input feature vector containing spatiotemporal coordinates, incoming flow conditions and wind turbine characteristic parameters, and acquire the real flow field physical quantity data corresponding to the input feature vector, which together constitute the training dataset.

[0107] The model building module 102 is used to build a deep neural network model, whose input is the input feature vector and whose output is the corresponding predicted value of the flow field physical quantity. Based on the fluid dynamics control equations followed by the wind turbine wake, the residuals of the physical equations are calculated using the output of the deep neural network and automatic differentiation technology. A composite loss function is constructed, including a data fitting loss term, a physical constraint loss term, a boundary condition loss term, and an initial condition loss term. The data fitting loss term is calculated based on the difference between the actual flow field physical quantity data and the predicted value of the flow field physical quantity.

[0108] The model training module 103 is used to train the deep neural network model by minimizing the composite loss function using the training dataset to obtain a wind turbine wake field prediction model.

[0109] Application module 104 is used to use the wind turbine wake field prediction model to simulate and predict the wake field of the target wind turbine.

[0110] It should be understood that the application of this application is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Those skilled in the art can understand that implementing all or part of the processes of the above embodiments and making equivalent changes according to the claims of this application still fall within the scope of this application.

Claims

1. A method for modeling and predicting the wake field of a wind turbine based on a physical information neural network, characterized in that, include: Multi-source observation data of the flow field where the wind turbine is located and wind turbine design parameters are obtained. An input feature vector containing spatiotemporal coordinates, incoming flow conditions and wind turbine characteristic parameters is constructed. The real flow field physical quantity data corresponding to the input feature vector are obtained and together they form a training dataset. A deep neural network model is constructed, with the input feature vector as its input and the corresponding predicted values ​​of flow field physical quantities as its output. Based on the fluid dynamics control equations followed by the wind turbine wake, the residuals of the physical equations are calculated using the output of the deep neural network and automatic differentiation techniques. A composite loss function is constructed, comprising a data fitting loss term, a physical constraint loss term, a boundary condition loss term, and an initial condition loss term. The data fitting loss term is calculated based on the difference between the actual flow field physical quantity data and the predicted values ​​of the flow field physical quantities. Using the training dataset, the deep neural network model is trained by minimizing the composite loss function to obtain a wind turbine wake field prediction model; The wind turbine wake field prediction model is used to simulate and predict the wake field of the target wind turbine.

2. The method for modeling and predicting the wake field of a wind turbine based on a physical information neural network according to claim 1, characterized in that, The input feature vector includes at least the spatial coordinates x, y, z, the time coordinate t, the incoming wind speed U∞, the wind direction θ, the wind turbine diameter D, the thrust coefficient Ct, and the power coefficient Cp.

3. The method for modeling and predicting the wake field of a wind turbine based on a physical information neural network according to claim 1, characterized in that, The deep neural network model includes two fully connected hidden layers connected in sequence, each containing 100 neurons, and the activation function is either the hyperbolic tangent function or the Swish function; the output layer of the deep neural network model directly outputs the predicted values ​​of the three-dimensional velocity components u, v, w and pressure p.

4. The method for modeling and predicting the wake field of a wind turbine based on a physical information neural network according to claim 1, characterized in that, The physical equation residuals include the continuity equation residuals R. cont With the residual R of the momentum equation x ,R y ,R z It automatically differentiates the partial derivatives of the predicted flow field physical quantities output by the deep neural network model and substitutes them into the following Navier-Stokes equations to obtain: Where u, v, w are the predicted values ​​of the fluid velocity components in the x, y, z directions output by the deep neural network model, p is the predicted value of the fluid pressure output by the deep neural network model, ρ is the fluid density, ν is the fluid kinematic viscosity, and fx, fy, fz are the volume force components used to simulate the force of the wind turbine blades on the airflow.

5. The method for modeling and predicting the wake field of a wind turbine based on a physical information neural network according to claim 1, characterized in that, The initial condition loss term The expression is: ； in, The total number of initial condition points, where each initial condition point is a spatiotemporal coordinate point with a known true flow field state at the initial time t=0; m is the index of the initial condition point; u m (t=0), v m (t=0) and w m (t=0) represents the velocity components of the fluid in the x, y and z directions predicted by the neural network model at the m-th initial condition point at the initial time t=0; and These are the true velocity components in the x, y, and z directions of the fluid at the m-th initial condition point recorded in the training dataset at the initial time t=0.

6. The method for modeling and predicting the wake field of a wind turbine based on a physical information neural network according to claim 1, characterized in that, The physical constraint loss term The expression is: ； Let i be the number of randomly sampled configuration points within the spatiotemporal computational domain of the wind turbine wake field, and let i be the index of the configuration point. For the continuity equation residual at the i-th location point, , and Let be the momentum equation residuals in the x, y, and z directions at the i-th placement point, respectively. for The weighting coefficients, , and They are respectively , and The weighting coefficients.

7. The method for modeling and predicting the wake field of a wind turbine based on a physical information neural network according to claim 1, characterized in that, The data fitting loss term The expression is: ； in, The total number of data points in the training dataset. , and These represent the velocity components of the fluid in the x, y, and z directions predicted by the neural network model at the j-th data point. This represents the fluid pressure predicted by the neural network model at j data points. , and Let x, y, and z represent the actual velocity components in the x, y, and z directions corresponding to the j-th data point, respectively. Let represent the actual pressure corresponding to the j-th data point.

8. The method for modeling and predicting the wake field of a wind turbine based on a physical information neural network according to claim 1, characterized in that, Boundary condition loss term The expression is: in, The total number of boundary condition points, where each boundary condition point is a spatiotemporal coordinate point with a known flow field state on the boundary of the computational domain of the wind turbine wake field; k is the index of the boundary condition point. , and These are the velocity components of the fluid in the x, y, and z directions predicted by the neural network model at the k-th boundary condition point. , and These are the x, y, and z-direction boundary velocity components of the fluid at the k-th boundary condition point, as preset according to the physical laws of the flow field.

9. The method for modeling and predicting the wake field of a wind turbine based on a physical information neural network according to claim 1, characterized in that, The steps of training the deep neural network model using the training dataset by minimizing the loss function include: The weights and bias parameters of the deep neural network model are randomly initialized and assigned values; Training batches are sampled from the training dataset, and multiple rounds of iterative optimization are performed. Each iteration includes: inputting the spatiotemporal coordinates of the training batches into the model for forward propagation to obtain predicted values ​​of flow field physical quantities; calculating the partial derivatives of the predicted values ​​using automatic differentiation; calculating the composite loss function based on the predicted values ​​and their partial derivatives, combined with the real physical quantity labels in the training dataset; calculating the gradient of the loss function with respect to the model parameters through backpropagation, and updating the model parameters using an adaptive moment estimation optimization algorithm or an L-BFGS optimization algorithm. During or after the training iteration optimization process, the model performance is evaluated using the validation dataset; hyperparameters are adjusted based on the validation results, including at least one of network depth, network width, activation function type, weight coefficients of each item in the composite loss function, learning rate, and batch size; the training iteration optimization steps are continued or re-executed based on the adjusted hyperparameters. When the preset convergence condition is met, training stops and the current model parameters are output to obtain the wind turbine wake field prediction model; the convergence condition is: the value of the composite loss function is lower than the preset threshold, or the rate of change of the composite loss function in multiple consecutive iterations is less than the stability threshold, or the preset maximum number of iterations is reached.

10. A device for modeling and predicting the wake field of a wind turbine based on a physical information neural network, characterized in that, include: The acquisition module is used to acquire multi-source observation data of the flow field where the wind turbine is located and the design parameters of the wind turbine, construct an input feature vector containing spatiotemporal coordinates, incoming flow conditions and wind turbine characteristic parameters, and acquire the real flow field physical quantity data corresponding to the input feature vector, which together constitute the training dataset. The model building module is used to construct a deep neural network model, whose input is the input feature vector and whose output is the corresponding predicted value of the flow field physical quantity. Based on the fluid dynamics control equations followed by the wind turbine wake, the residuals of the physical equations are calculated using the output of the deep neural network and automatic differentiation techniques. A composite loss function is constructed, including a data fitting loss term, a physical constraint loss term, a boundary condition loss term, and an initial condition loss term. The data fitting loss term is calculated based on the difference between the actual flow field physical quantity data and the predicted value of the flow field physical quantity. The model training module is used to train the deep neural network model by minimizing the composite loss function using the training dataset to obtain a wind turbine wake field prediction model. The application module is used to use the wind turbine wake field prediction model to simulate and predict the wake field of the target wind turbine.

Images