A method and system for predicting the velocity and turbulence intensity of the wake field of a tidal current energy unit based on physical consistency constraints

Abstract

This invention discloses a method and system for predicting wake field velocity and turbulence intensity of tidal power units based on physical consistency constraints. The method includes first constructing a dataset through simulation, then building a wake field velocity and turbulence intensity prediction model, inputting the dataset into the prediction model for training, obtaining a trained wake field velocity and turbulence intensity prediction model, deploying the trained model under actual operating conditions, and collecting the wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit in real time and inputting them into the prediction model to predict the corresponding wake field velocity and turbulence intensity. This invention can achieve high-precision wake field prediction while ensuring computational efficiency, and is particularly suitable for wake field prediction under complex operating conditions and optimization of tidal power unit array layout.

Description

Technical Field

[0001] This invention belongs to the fields of tidal power generation, artificial intelligence and fluid mechanics, and specifically relates to a method and system for predicting the velocity and turbulence intensity of the wake field of a tidal power unit based on physical consistency constraints. Background Technology

[0002] Tidal energy, as an important renewable energy source, has received widespread attention in recent years due to the increasing global demand for clean energy. Tidal turbines, which convert the kinetic energy of ocean currents into electricity, play a crucial role in ocean energy development. However, when arranged in an array, the performance of tidal turbines is significantly affected by the wake effect. The low-speed eddies in the wake region not only cause power losses in downstream tidal turbines but also exacerbate the unsteady loads on the turbines, thus impacting the overall array layout optimization and energy output assessment.

[0003] Currently, wake prediction methods for tidal power units mainly rely on analytical models and numerical simulations. While analytical models are computationally efficient, they have limitations in accuracy and generalization ability, especially in accurately capturing turbulence and vortex characteristics under complex flow fields and variable operating conditions. Numerical simulations (such as CFD methods) offer higher prediction accuracy, but due to their massive computational demands, they are insufficient to meet the needs of large-scale array layout optimization and real-time operational evaluation.

[0004] Patent CN202311728507.4 discloses a method for predicting wind farm wake and power based on a convolutional neural network model. This method uses datasets generated by analytical wake models or numerical models to predict the wake field and power of wind turbines. However, this method has drawbacks: it relies heavily on the accuracy of simulation and measured data, making it unable to accurately handle complex turbulent and vortex structures, and its prediction accuracy is limited, especially under varying flow field conditions.

[0005] Patent CN202311734604.4 discloses a method for predicting wind farm wake and power based on a generative adversarial network model. This method utilizes a conditional generative adversarial network (CGAN) based on a Transformer model to construct a single wind turbine wake prediction model. However, this method has drawbacks: GAN networks are prone to mode collapse (inconsistent results) during training, and require a large training dataset. While this dataset may be sufficient for wind turbines, it may be insufficient for tidal power turbines not yet commercially operational, limiting model stability. Furthermore, although the method integrates generative adversarial networks and Transformer models, it does not explicitly incorporate constraints related to physical consistency (such as momentum equations and turbulence closure), which may lead to prediction results lacking physical plausibility.

[0006] Patent CN202511339876.3 discloses a method and device for predicting the flow field of a wind turbine under multiple operating conditions based on a neural network, trained using a Domain Adversarial Neural Network (DANN) model. The drawbacks of this method are: although it introduces a domain adversarial training mechanism, it still heavily relies on diverse training data. Furthermore, due to the different flow field characteristics of tidal energy, this method may not be directly applicable to specific scenarios such as tidal power units, and additional adaptation may be required. Summary of the Invention

[0007] To address the problems existing in the background technology, this invention proposes a method and system for predicting the wake velocity and turbulence intensity of tidal power units based on physical consistency constraints, which solves the technical problems of lack of physical consistency, insufficient generalization ability, and low accuracy and efficiency in the prior art.

[0008] This invention proposes a physical consistency constraint loss function and a dynamic weighting strategy to construct a model that combines physical constraints with data-driven approaches. This model can achieve high-precision wake field prediction while ensuring computational efficiency. This invention is particularly suitable for wake field prediction under complex operating conditions and for optimizing the layout of tidal power unit arrays.

[0009] The technical solution adopted in this invention is: I. A method for predicting the velocity and turbulence intensity in the wake field of a tidal power unit based on physical consistency constraints: S1. The wake field model is built using the wall adaptive local eddy viscosity model in the large eddy simulation method, and the wake field data is obtained through simulation. A dataset is constructed based on the wake field data.

[0010] The wake field data includes the wake field coordinates, inflow velocity, inflow turbulence intensity, and corresponding wake field velocity and turbulence intensity of the tidal power unit. The wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit are used as input subsamples, and the corresponding wake field velocity and turbulence intensity are used as output subsamples to form a single sample pair of the dataset.

[0011] S2. Construct a wake field velocity and turbulence intensity prediction model. Input the dataset into the wake field velocity and turbulence intensity prediction model for training to obtain the trained wake field velocity and turbulence intensity prediction model.

[0012] S3. Deploy the trained wake field velocity and turbulence intensity prediction model to the actual operating conditions, collect the wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit in real time, and input them into the prediction model to predict the corresponding wake field velocity and turbulence intensity.

[0013] The wake field velocity and turbulence intensity prediction model includes a preprocessing module and a prediction module connected in series. The input of the preprocessing module serves as the input of the wake field velocity and turbulence intensity prediction model, and the output of the prediction module serves as the output of the wake field velocity and turbulence intensity prediction model. The preprocessing module performs data standardization, missing value processing, and noise reduction on the input data in sequence. The prediction module employs a multilayer perceptron.

[0014] The total loss function of the prediction module is set according to the following formula: ; ; ; in, The total loss function for the prediction module; For data loss function; The loss function is the physical consistency constraint. These are the weighting coefficients of the physical consistency constraint loss function; For indexing; This represents the total number of training samples in a single round. For the first Simulated wake velocity values ​​for each sample; For the first Predicted wake velocity values ​​for each sample; and All are weighting coefficients; For the first Simulated turbulence intensity values ​​for each sample; For the first Predicted turbulence intensity values ​​for each sample; Loss due to boundary consistency constraints; The loss function is the momentum convection-diffusion residual constraint. This is the loss due to turbulent closure residual constraints. , and All are weighting coefficients; For indexing; This represents the number of points after boundary sampling of the wake field; and All are weighting coefficients; For the first Predicted wake velocity values ​​at each boundary sampling point; For the first The corresponding inflow velocity at each boundary sampling point; For the first Predicted turbulence intensity at each boundary sampling point; For the first The corresponding inflow turbulence intensity at each boundary sampling point; , and All are indexes; The number of points after physical sampling of the wake field; For the first The planar coordinates of each physical sampling point; plane coordinates First Large eddy simulation momentum residuals in each direction; The fluid's kinematic viscosity coefficient; The eddy viscosity coefficient is at the subgrid scale. and plane coordinates First Components of the directional wake field velocity and the Components of the wake field velocity in each direction The values ​​after filtering; plane coordinates First Coordinate components in each direction; For sublattice-scale stress tensor; It is a first-order partial derivative; It is a second-order partial derivative; subgrid-scale stress tensor traces; In plane coordinates Turbulent closure residuals at the location; Kronecker symbol; Here is the eddy viscosity coefficient under the wall adaptive local eddy viscosity model; plane coordinates First Components of the directional wake field velocity and the Components of the wake field velocity in each direction The product is the value that is then filtered.

[0015] The prediction module employs a dynamic weighting strategy to adjust weights during training; the weights of each neuron in the prediction module are adjusted according to the following formula of the dynamic weighting strategy: in, For the adjusted current time The weights of the lower neurons; The coefficient of the exponential moving average; To adjust the previous moment The weights of the lower neurons; For the clip function; For data loss function; The loss function is the physical consistency constraint. To prevent division by zero of constants; To adjust the previous moment The minimum value of the weights of all neurons; To adjust the previous moment The maximum value of all neuron weights.

[0016] S4. Obtain the average inflow velocity of the tidal power unit at the impeller section based on the real-time predicted wake field velocity, and obtain the real-time power of the tidal power unit based on the average inflow velocity.

[0017] The power of the tidal current generator unit is obtained according to the following formula: ; in, The power of the tidal current generator unit; C p Power coefficient; ρ The density of water; For the efficiency of tidal power units; D The impeller diameter of the tidal power unit; The average inflow velocity at the impeller cross-section; This represents a small area element within the swept area of ​​the impeller of a tidal power unit. This represents the average wake velocity within the impeller swept region corresponding to the impeller cross-section of the tidal power unit.

[0018] II. A prediction system for wake field velocity and turbulence intensity of tidal power units based on physical consistency constraints: The dataset construction module uses the wall adaptive local eddy viscosity model in the large eddy simulation method to build a wake field model and simulate the wake field data, and then constructs a dataset based on the wake field data.

[0019] The wake field velocity and turbulence intensity prediction model training module constructs a wake field velocity and turbulence intensity prediction model. The dataset is input into the wake field velocity and turbulence intensity prediction model for training, and the trained wake field velocity and turbulence intensity prediction model is obtained.

[0020] The online inference module deploys the trained wake field velocity and turbulence intensity prediction model to actual operating conditions, and collects the wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit in real time and inputs them into the prediction model to predict the corresponding wake field velocity and turbulence intensity.

[0021] The beneficial effects of this invention are: 1) Improving Prediction Accuracy and Computational Efficiency by Generating High-Fidelity Datasets Using Numerical Models: This invention utilizes the wall-adaptive local eddy viscosity model from Large Eddy Simulation (LES) to build a wake field model, thereby providing more accurate training data. These datasets can better capture turbulence and vortex structures in complex flow fields, improving the prediction accuracy of wake field velocity and turbulence intensity. Compared with traditional analytical wake models, the datasets generated by the wall-adaptive local eddy viscosity model provide more realistic and high-fidelity training data for deep learning models, resulting in higher prediction accuracy under different operating conditions, while avoiding the high computational cost of numerical simulation, thus improving computational efficiency.

[0022] 2) Adaptability across operating conditions: By explicitly introducing operating condition features such as inflow velocity and turbulence intensity into the input and training the model with a dataset covering multiple operating conditions, this invention enables the model to possess powerful cross-operating condition prediction capabilities. Employing a dynamic weighting and exponential moving average mechanism, the model can adaptively balance data loss and physical loss, and optimize hyperparameters through Bayesian optimization, thereby improving training stability, prediction accuracy, and generalization ability, effectively addressing variations in the operating conditions of tidal power units.

[0023] 3) Ensuring the physical rationality and reliability of prediction results: This invention introduces physical consistency constraints such as boundary consistency, momentum convection-diffusion residuals, and turbulence closure residuals during the training process. These constraints are incorporated into the loss function and optimized together with the data loss, thereby ensuring the physical rationality of the model's prediction results. Simultaneously, through a bounded output mapping mechanism, the generation of non-physical results is avoided, significantly improving the reliability of the prediction results and ensuring stability and accuracy under multiple operating conditions, providing reliable data support for the performance evaluation of tidal power units. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention.

[0025] Figure 2 This is a schematic diagram of the sampling point setup for generating the dataset using the large eddy simulation method in this embodiment.

[0026] Figure 3 This is a schematic diagram of the structure of the multilayer sensor of the present invention.

[0027] Figure 4 Isv in =1.5m / s, TI in A schematic diagram showing the predicted results of wake field velocity and turbulence intensity under the 1% operating condition.

[0028] Figure 5 Is v in =2m / s, TI in A schematic diagram showing the predicted results of wake field velocity and turbulence intensity under the 5% operating condition.

[0029] Figure 6 Is v in =2.5m / s, TI in A schematic diagram showing the predicted results of wake field velocity and turbulence intensity under the 15% operating condition. Detailed Implementation

[0030] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0031] like Figure 1 and Figure 2 As shown, the method for predicting the wake velocity and turbulence intensity of a tidal power unit in this embodiment is implemented according to the following steps: S1. The wake field model was built and the wake field data were obtained by simulation using the Wall-Adapting Local Eddy-viscosity (WALE) model in the Large Eddy Simulation (LES) method under multiple different incoming flow velocities and turbulence intensities. A dataset was constructed based on the wake field data.

[0032] The wake field data includes the wake field coordinates, inflow velocity, inflow turbulence intensity, and corresponding wake field velocity and turbulence intensity of the tidal power unit. The wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit are used as input subsamples, and the corresponding wake field velocity and turbulence intensity are used as output subsamples to form a single sample pair of the dataset.

[0033] S2. Construct a wake field velocity and turbulence intensity prediction model. Input the dataset into the wake field velocity and turbulence intensity prediction model for training to obtain the trained wake field velocity and turbulence intensity prediction model.

[0034] The wake field velocity and turbulence intensity prediction model consists of a preprocessing module and a prediction module connected in series. The input of the preprocessing module serves as the input to the wake field velocity and turbulence intensity prediction model, and the output of the prediction module serves as the output. The preprocessing module performs data standardization, missing value handling, and noise reduction on the input data sequentially.

[0035] Furthermore, the preprocessing module performs comprehensive preprocessing on the data to ensure it meets the training requirements of the neural network. First, data standardization is used to normalize the input subsamples (such as inflow velocity and inflow turbulence intensity) so that their distribution meets the input requirements of the neural network. Next, missing values ​​in the dataset are filled or deleted to ensure data integrity and accuracy. In addition, smoothing techniques (such as mean filtering or Gaussian filtering) are used to remove abnormal noise from the data, improving data quality and ensuring stability and accuracy during training. Through these steps, the data is optimized, thus providing high-quality input for subsequent model training.

[0036] Specifically, 1) Data standardization: Standardize the input features such as inflow velocity and turbulence intensity to make the data evenly distributed within a specified range (such as the interval between 0 and 1), thereby improving the stability of the neural network training process.

[0037] 2) Missing value handling: Use interpolation or mean imputation to handle missing values ​​in the dataset to ensure that each sample does not lose key information during training.

[0038] 3) Noise removal: Apply smoothing techniques (such as Gaussian filtering or local weighted regression) to remove noise from the data and reduce errors caused by outliers during model learning.

[0039] The prediction module employs a multilayer perceptron. For example... Figure 3 As shown, the multilayer perceptron in this embodiment consists of several hidden layers and neurons (located on each hidden layer). The number of hidden layers and the number of neurons in each layer are automatically determined through hyperparameter optimization, and the output layer uses the Sigmoid function for bounded output mapping to ensure the physical feasibility of the prediction results.

[0040] In this embodiment, in order to ensure that the neural network prediction results can meet the physical laws of wake evolution, this embodiment introduces multiple physical consistency constraints during the training process, which are used as part of the loss function and optimized together with the data loss to ensure the physical rationality and consistency of the prediction results.

[0041] Therefore, the total loss function of the prediction module is set according to the following formula: in, The total loss function for the prediction module; For data loss function; The loss function is the physical consistency constraint. These are the weighting coefficients of the physical consistency constraint loss function. in, For indexing; This represents the total number of training samples in a single round. For the first Simulated wake velocity values ​​for each sample; For the first Predicted wake velocity values ​​for each sample; and All are weighting coefficients; For the first Simulated turbulence intensity values ​​for each sample; For the first Predicted turbulence intensity values ​​for each sample. in, Loss due to boundary consistency constraints; The loss function is the momentum convection-diffusion residual constraint. This is the loss due to turbulent closure residual constraints. , and All are weighting coefficients.

[0042] Boundary consistency constraints (BC) are set at the boundaries of the physical sampling points (the equivalent boundary zones of the upper and lower boundaries of the wake field or the inlet boundary) to ensure that the predicted velocity and turbulence intensity of the wake field are consistent with the inflow conditions. The principle behind this is that the wake field of a tidal power unit is an open flow system controlled by the incoming flow conditions. The flow state near the inlet boundary and the upper and lower boundaries should be consistent with or continuously transition with the given inflow velocity and inflow turbulence intensity. If the predicted results at the boundaries deviate from the inflow conditions, non-physical velocity abrupt changes or turbulence intensity anomalies can easily occur at the outer edge of the wake field, further affecting the prediction accuracy within the wake field. Therefore, this embodiment applies consistency constraints at the boundary sampling points to the predicted wake field velocity and the corresponding inflow velocity, and the predicted turbulence intensity and the corresponding inflow turbulence intensity. By calculating the boundary loss and using mean squared error (MSE) to optimize the boundary data, the boundary conditions are explicitly embedded as prior physical information into the model training process, ensuring the physical consistency of the model predictions. By introducing this constraint, non-physical oscillations, abnormal deviations, and discontinuities in the boundary region can be effectively suppressed, the propagation of boundary errors into the wake field can be reduced, the stability of model training and the generalization ability under different inflow velocities and turbulence intensities can be improved, thereby enhancing the physical rationality and reliability of the wake field velocity and turbulence intensity prediction results. in, For indexing; This represents the number of points after boundary sampling of the wake field; and All are weighting coefficients; For the first Predicted wake velocity values ​​at each boundary sampling point; For the first The corresponding inflow velocity at each boundary sampling point; For the first Predicted turbulence intensity at each boundary sampling point; For the first The corresponding inflow turbulence intensity at each boundary sampling point.

[0043] To ensure that the predicted wake field velocity satisfies the constraints of the large eddy simulation filter momentum equation, this embodiment uses physical sampling points within the wake region of the tidal power unit. A momentum convection-diffusion residual is constructed, and its mean square error is used as part of the loss function in network training. The underlying principle is that the formation and evolution of the wake field of a tidal power unit are essentially controlled by the fluid momentum transport process. The propagation of velocity deficit, the development of the shear layer, and the wake recovery process within the wake region are all related to the convection term, the molecular viscous diffusion term, and the subgrid-scale stress term. If only sample data is used to fit the wake field velocity, although small errors can be obtained at some sampling locations, when the training sample is insufficient, the operating conditions vary greatly, or the wake structure is complex, it is easy to encounter situations where the local velocity distribution is approximately consistent with the sample, but the overall momentum conservation relationship and wake evolution law are not satisfied. Based on this, in this embodiment, at the physical sampling points inside the wake region, the filtered velocity field predicted by the neural network is substituted into the large eddy simulation filtering momentum equation to construct the corresponding momentum convection-diffusion residual. By constraining the residual to tend to decrease, the prediction result satisfies the basic physical laws of axial momentum transfer, lateral momentum diffusion and subgrid-scale momentum exchange while fitting the data. Thus, the momentum evolution mechanism inside the wake is explicitly embedded into the model training process.

[0044] Momentum convection-diffusion residual constraint: By controlling the axial momentum transport characteristics of the momentum equation residual model through large eddy simulation (LES), the predicted wake velocity is consistent with the laws of fluid dynamics. This constraint essentially transforms the traditional numerical solution of discretely solving the control equation into a soft constraint term in neural network training. By penalizing the momentum equation residual at physical sampling points, it suppresses non-physical aspects in the predicted velocity field that do not conform to the laws of wake convection, diffusion, and stress transfer. Furthermore, since the LES framework used in this embodiment has characterized the subgrid-scale effect through a wall-adaptive local eddy viscosity model, this residual constraint can maintain consistency with the subsequent turbulence closure residual constraint under the same physical modeling basis. This ensures that the predicted wake velocity not only numerically approximates the sample data but also conforms to the laws of filtered momentum transport in terms of physical mechanisms. By introducing this constraint, non-physical velocity oscillations and anomalous gradients in the wake center region, shear layer region, and wake recovery region can be effectively reduced. This avoids local distortion or momentum transport imbalance in model predictions across different operating conditions, improves the stability of model training convergence, its ability to represent complex flow structures, and its generalization ability under different inflow velocities and turbulence intensities. Consequently, it enhances the physical rationality, accuracy, and engineering reliability of wake field velocity prediction results. ; in, , and All are indexes; The number of points after physical sampling of the wake field; For the first The planar coordinates of each physical sampling point (i.e., coordinate values ​​in two directions); plane coordinates First Large eddy simulation momentum residuals in each direction; The fluid's kinematic viscosity coefficient; The eddy viscosity coefficient is at the subgrid scale. and plane coordinates First Components of the directional wake field velocity and the Components of the wake field velocity in each direction The values ​​after filtering; plane coordinates First Coordinate components in each direction; The stress tensor is at the subgrid scale (determined based on the LES-WALE adaptive local eddy viscosity model in the large eddy simulation method). It is a first-order partial derivative; It is a second-order partial derivative.

[0045] To ensure the consistency of subgrid-scale (SGS) stress closure in Large Eddy Simulation (LES), this embodiment employs a wall-adaptive local eddy viscosity model to close the eddy viscosity coefficient. Furthermore, it constructs turbulent closure residual constraints to maintain physical consistency, thus ensuring consistency with the momentum convection-diffusion residual constraints within the same LES framework. The underlying principle is that in LES, the large-scale flow structure of the wake field can be directly analyzed by the filtered momentum equation, while the influence of small-scale turbulent fluctuations on the large-scale flow is reflected through the subgrid-scale stress term. Therefore, the closure method of the subgrid-scale stress directly affects the characterization accuracy of wake shear layer development, turbulent diffusion, and wake recovery processes. If only wake field velocity and turbulence intensity are fitted without constraining the closure relationship between subgrid-scale stress and eddy viscosity coefficient, the model is prone to producing predictions that, while having small errors at local data points, are inconsistent with actual turbulent transport mechanisms in terms of physical mechanisms. This leads to distortion of turbulence intensity distribution and deviation of momentum transport relationships. Based on this, this embodiment uses a wall-adaptive local eddy viscosity model to characterize the eddy viscosity coefficient, and constructs a turbulence closure residual at the physical sampling point based on the predicted turbulence intensity and related velocity information, so that the subgrid-scale stress tensor, eddy viscosity coefficient and local flow state satisfy a consistent closure relationship, thereby explicitly embedding the turbulence subgrid-scale modeling mechanism into the neural network training process.

[0046] The turbulence closure residual constraint, based on the predicted turbulence intensity, combines the wall-adaptive local eddy viscosity model to close the eddy viscosity coefficient, and constructs a turbulence closure residual constraint to ensure its physical consistency, thereby improving the accuracy of turbulence intensity prediction. This constraint essentially transforms the closure relationship of the subgrid-scale model in traditional large eddy simulation into a physical constraint term in neural network training. By penalizing the closure residual, it suppresses non-physical interpretations in the prediction results that do not conform to local shear characteristics, stress response laws, and turbulence dissipation characteristics. Furthermore, since the wall-adaptive local eddy viscosity model can better characterize the local eddy viscosity effect in the near-wall region and the strong shear region, the turbulence closure residual constraint constructed based on the wall-adaptive local eddy viscosity model in this embodiment can more effectively reflect the coupling relationship between the evolution of turbulence intensity and subgrid-scale stress in the wake region of the tidal power unit. Combined with the aforementioned momentum convection-diffusion residual constraint, the predicted wake field velocity and turbulence intensity not only match the sample data but also satisfy the unified large eddy simulation physical framework. By introducing this constraint, we can effectively reduce the abnormally high and low predicted turbulence intensity and local oscillations in the wake shear layer, wake center region, and wake recovery region. This improves the model's ability to represent complex turbulent structures and cross-condition flow characteristics, and enhances the training convergence stability, turbulence intensity prediction accuracy, and the physical rationality and engineering reliability of the overall prediction results. ; ; subgrid-scale stress tensor traces; In plane coordinates Turbulent closure residuals at the location; Kronecker symbol; Here is the eddy viscosity coefficient under the wall adaptive local eddy viscosity model; plane coordinates First Components of the directional wake field velocity and the Components of the wake field velocity in each direction The product is the value that is then filtered.

[0047] In practice , and ,Right now When choosing the first direction, Choose the second direction; When choosing the second direction, Choose the first direction.

[0048] like Figure 2As shown, when obtaining the physical consistency constraint loss function, a physical sampling strategy is used for sampling, that is, a certain number of physical points are sampled in the wake field in each training cycle: For physical points Physical point The first coordinate value Uniform sampling is performed within the coordinate range of the training set, for physical points. The second coordinate value Small perturbations are sampled and superimposed near the upper and lower boundaries; simultaneously, physical points are randomly selected from the training set. Inflow velocity for corresponding operating conditions v in and inflow turbulence intensity TI in As this physical point The operating conditions are used as input to achieve physical constraints across operating conditions.

[0049] Figure 2 In this context, D represents the impeller diameter of the tidal power unit. This indicates that the wake field calculation domain is set to a cylindrical region with a diameter of 100m, that is, this cylindrical region is used as the spatial range for numerical calculation and prediction analysis of the wake field.

[0050] In this embodiment, at the inflow velocity v in =1.5m / s, inflow turbulence intensity TI in =1% operating condition, each physical point Simulated wake field velocity The corresponding predicted wake field velocity and the corresponding relative error like Figure 4 .

[0051] Inflow velocity v in =2m / s, inflow turbulence intensity TI in =5% operating condition, each physical point Simulated wake field velocity The corresponding predicted wake field velocity and the corresponding relative error like Figure 5 .

[0052] Inflow velocity v in =2.5m / s, inflow turbulence intensity TI in =15% operating condition, each physical point Simulated wake field velocity The corresponding predicted wake field velocity and the corresponding relative error like Figure 6 .

[0053] The training process and loss function optimization are achieved through the following steps: 1) Dynamic weighting strategy: By dynamically adjusting the weight ratio between physical loss and data loss, the loss function is optimized by using an exponential moving average mechanism to improve the stability and accuracy of training.

[0054] 2) Hyperparameter Optimization: An automated hyperparameter optimization framework is introduced. Through Bayesian optimization, parameters in the network architecture, learning rate, physical loss weights, and dynamic weighting strategy are optimized, automatically selecting the optimal hyperparameter configuration. The hyperparameter optimization framework can adaptively adjust parameters based on the model's performance on the validation set, avoiding the complexity of manual parameter tuning and achieving the best training results under different conditions.

[0055] 3) Intelligent adjustment: During the training process, the weighting ratio of the loss function and the neural network structure are continuously optimized through hyperparameter optimization to ensure that the model can adapt to different input data and complex training conditions, thereby improving the model's prediction accuracy and generalization ability.

[0056] The prediction module employs a dynamic weighting strategy for weight adjustment during training; the weights of each neuron in the prediction module are adjusted according to the following formula: in, For the adjusted current time The weights of the lower neurons; The coefficient of the exponential moving average; To adjust the previous moment The weights of the lower neurons; For the clip function; For data loss function; The loss function is the physical consistency constraint. To prevent division by zero of extremely small constants; To adjust the previous moment The minimum value of the weights of all neurons; To adjust the previous moment The maximum value of all neuron weights.

[0057] Specifically, Used to limit the range of weights, used to assign input values Limited to Within the interval, if Less than Then return If greater than Then return .

[0058] Specifically, the AdamW optimizer is used to iteratively update network parameters, and learning rate warm-up and cosine annealing strategies can be set to improve convergence stability; early stopping is performed based on the validation set data loss, and training is stopped and the optimal model is saved when the validation set loss no longer decreases within a preset number of rounds.

[0059] S3. Deploy the trained wake field velocity and turbulence intensity prediction model to the actual operating conditions, collect the wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit in real time, and input them into the prediction model to predict the corresponding wake field velocity and turbulence intensity.

[0060] S4. Obtain the average inflow velocity of the tidal power unit at the impeller section based on the real-time predicted wake field velocity, and obtain the real-time power of the tidal power unit based on the average inflow velocity.

[0061] Specifically, the method for obtaining the average inflow velocity of the tidal power unit at the impeller section based on the wake field velocity predicted in real time is as follows: Using the impeller cross-section at the location of the tidal power unit where the power to be determined is located as the statistical cross-section, the average wake velocity in the impeller swept region corresponding to the impeller cross-section is obtained. Thus, the average inflow velocity at the impeller section of the tidal power unit is obtained. . in, The average inflow velocity at the impeller cross-section; D The impeller diameter of the tidal power unit; This represents a small area element within the swept area of ​​the impeller of a tidal power unit. This represents the average wake velocity within the impeller swept region corresponding to the impeller cross-section of the tidal power unit.

[0062] In other words, the average inflow velocity at the impeller cross-section To predict the average area of ​​the wake field velocity on the corresponding impeller swept surface This reflects the equivalent inflow velocity actually acting on the impeller of the tidal power unit after being affected by the upstream wake. Based on this average inflow velocity... Furthermore, by combining the power coefficient, water density, impeller swept area, and overall efficiency of the tidal power unit, the real-time power of the tidal power unit can be calculated. Therefore, the power of the tidal power unit, as a derived quantity further derived from the predicted wake field velocity, can establish a direct correlation between the wake field prediction results and the unit power assessment, thereby achieving real-time prediction and assessment of the operating performance of the tidal power unit.

[0063] The power output of the tidal power unit is obtained using the following formula: in, The power of the tidal current generator unit; C p Power coefficient; ρ The density of water; For the efficiency of the tidal power unit.

[0064] Specifically, the average inflow velocity of the tidal power unit at the impeller section is obtained based on the wake field velocity predicted in real time.

[0065] Furthermore, error analysis is performed between the predicted real-time power of the tidal power unit and the simulation results to compare the performance of the model in terms of power output.

[0066] This embodiment also provides a system for predicting the wake velocity and turbulence intensity of a tidal power unit, including: The dataset construction module uses the wall adaptive local eddy viscosity model in the large eddy simulation method to build a wake field model and simulate the wake field data, and then constructs a dataset based on the wake field data.

[0067] The wake field velocity and turbulence intensity prediction model training module constructs a wake field velocity and turbulence intensity prediction model. The dataset is input into the wake field velocity and turbulence intensity prediction model for training, and the trained wake field velocity and turbulence intensity prediction model is obtained.

[0068] The online inference module deploys the trained wake field velocity and turbulence intensity prediction model to actual operating conditions, and collects the wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit in real time and inputs them into the prediction model to predict the corresponding wake field velocity and turbulence intensity.

[0069] To demonstrate the beneficial effects of the total loss function proposed in this invention, this embodiment compares the prediction module with different existing models, and the comparison results are as follows: Table 1. Comparison of prediction performance among different models: To demonstrate the beneficial effects of the dynamic weighting strategy proposed in this invention, this embodiment conducted a comparative ablation experiment, and the comparison results are as follows: Table 2. Comparison of prediction performance with and without dynamic weighting strategy: This invention can achieve high-precision wake field prediction while ensuring computational efficiency, and is especially suitable for wake field prediction and tidal power unit array layout optimization under complex operating conditions.

[0070] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for predicting the velocity and turbulence intensity of the wake field of a tidal power unit based on physical consistency constraints, characterized in that, Includes the following steps: S1. The wake field model is built using the wall adaptive local eddy viscosity model in the large eddy simulation method, and the wake field data is obtained through simulation. A dataset is constructed based on the wake field data. The wake field data includes the wake field coordinates, inflow velocity, inflow turbulence intensity, and corresponding wake field velocity and turbulence intensity of the tidal power unit; the wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit are used as input subsamples, and the corresponding wake field velocity and turbulence intensity are used as output subsamples to form a single sample pair of the dataset; S2. Construct a prediction model for wake field velocity and turbulence intensity. Input the dataset into the prediction model for wake field velocity and turbulence intensity for training to obtain the trained prediction model for wake field velocity and turbulence intensity. S3. Deploy the trained wake field velocity and turbulence intensity prediction model to the actual operating conditions, collect the wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit in real time, and input them into the prediction model to predict the corresponding wake field velocity and turbulence intensity.

2. The method for predicting the wake velocity and turbulence intensity of a tidal power unit according to claim 1, characterized in that: The wake field velocity and turbulence intensity prediction model includes a preprocessing module and a prediction module connected in series. The input of the preprocessing module serves as the input of the wake field velocity and turbulence intensity prediction model, and the output of the prediction module serves as the output of the wake field velocity and turbulence intensity prediction model. The preprocessing module performs data standardization, missing value processing, and noise reduction on the input data in sequence. The prediction module employs a multilayer perceptron.

3. The method for predicting the wake velocity and turbulence intensity of a tidal power unit according to claim 2, characterized in that: The total loss function of the prediction module is set according to the following formula: ； ； ； in, The total loss function for the prediction module; For data loss function; The loss function is the physical consistency constraint. These are the weighting coefficients of the physical consistency constraint loss function; For indexing; This represents the total number of training samples in a single round. For the first Simulated wake velocity values ​​for each sample; For the first Predicted wake velocity values ​​for each sample; and All are weighting coefficients; For the first Simulated turbulence intensity values ​​for each sample; For the first Predicted turbulence intensity values ​​for each sample; Loss due to boundary consistency constraints; The loss function is the momentum convection-diffusion residual constraint. This is the loss due to turbulent closure residual constraints. , and All are weighting coefficients; For indexing; This represents the number of points after boundary sampling of the wake field; and All are weighting coefficients; For the first Predicted wake velocity values ​​at each boundary sampling point; For the first The corresponding inflow velocity at each boundary sampling point; For the first Predicted turbulence intensity at each boundary sampling point; For the first The corresponding inflow turbulence intensity at each boundary sampling point; , and All are indexes; The number of points after physical sampling of the wake field; For the first The planar coordinates of each physical sampling point; plane coordinates First Large eddy simulation momentum residuals in each direction; The fluid's kinematic viscosity coefficient; The eddy viscosity coefficient is at the subgrid scale. and plane coordinates First Components of the directional wake field velocity and the Components of the wake field velocity in each direction The values ​​after filtering; plane coordinates First Coordinate components in each direction; For sublattice-scale stress tensor; It is a first-order partial derivative; It is a second-order partial derivative; subgrid-scale stress tensor traces; In plane coordinates Turbulent closure residuals at the location; Kronecker symbol; Here is the eddy viscosity coefficient under the wall adaptive local eddy viscosity model; plane coordinates First Components of the directional wake field velocity and the Components of the wake field velocity in each direction The product is the value that is then filtered.

4. The method for predicting the wake velocity and turbulence intensity of a tidal power unit according to claim 2, characterized in that: The prediction module employs a dynamic weighting strategy to adjust weights during training; the weights of each neuron in the prediction module are adjusted according to the following formula of the dynamic weighting strategy: in, For the adjusted current time The weights of the lower neurons; The coefficient of the exponential moving average; To adjust the previous moment The weights of the lower neurons; For the clip function; For data loss function; The loss function is the physical consistency constraint. To prevent division by zero of constants; To adjust the previous moment The minimum value of the weights of all neurons; To adjust the previous moment The maximum value of all neuron weights.

5. The method for predicting the wake velocity and turbulence intensity of a tidal power unit according to claim 1, characterized in that, Also includes: S4. Obtain the average inflow velocity of the tidal power unit at the impeller section based on the real-time predicted wake field velocity, and obtain the real-time power of the tidal power unit based on the average inflow velocity.

6. The method for predicting the wake velocity and turbulence intensity of a tidal power unit according to claim 5, characterized in that: The power of the tidal current generator unit is obtained according to the following formula: ； in, The power of the tidal current generator unit; C p Power coefficient; ρ The density of water; For the efficiency of tidal power units; D The impeller diameter of the tidal power unit; The average inflow velocity at the impeller cross-section; This represents a small area element within the swept area of ​​the impeller of a tidal power unit. This represents the average wake velocity within the impeller swept region corresponding to the impeller cross-section of the tidal power unit.

7. A tidal power unit wake field velocity and turbulence intensity prediction system using the method described in any one of claims 1-6, characterized in that, include: The dataset construction module uses the wall adaptive local eddy viscosity model in the large eddy simulation method to build a wake field model and simulate the wake field data, and then constructs a dataset based on the wake field data. The wake field velocity and turbulence intensity prediction model training module constructs a wake field velocity and turbulence intensity prediction model. The dataset is input into the wake field velocity and turbulence intensity prediction model for training, and the trained wake field velocity and turbulence intensity prediction model is obtained. The online inference module deploys the trained wake field velocity and turbulence intensity prediction model to actual operating conditions, and collects the wake field coordinates, inflow velocity, and inflow turbulence intensity of the tidal power unit in real time and inputs them into the prediction model to predict the corresponding wake field velocity and turbulence intensity.

Images