Wind turbine three-dimensional space-time wind field prediction method and system based on frequency domain physical information neural network and storage medium

Abstract

The invention discloses a wind turbine three-dimensional space-time wind field prediction method based on a frequency domain physical information neural network. The wind turbine three-dimensional space-time wind field prediction method comprises the following steps that 1, wind speed data are collected through sensors of a plurality of measuring points arranged at the position of a wind turbine; 2, carrying out the data preprocessing of the wind speed data, selecting the wind speed data of a part of measurement points as a training set, and taking the wind speed data of the remaining measurement points as a verification set; 3, determining a physical model and frequency domain information of the three-dimensional space-time wind field according to the wind field data of the training set; 4, fusing the physical model and the frequency domain information of the three-dimensional space-time wind field into a neural network to construct an FD-PINN model; step 5, using the training set to train the FD-PINN model; using the verification set to evaluate the generalization ability of the trained FD-PINN model; and 6, collecting wind speed data in real time, and predicting to obtain a three-dimensional space-time wind field of the wind turbine. The invention further discloses a wind turbine three-dimensional space-time wind field prediction system based on the frequency domain physical information neural network and a storage medium.

Description

Technical Field

[0001] The present invention belongs to the field of new energy wind power generation and wind turbine structure safety assessment, and specifically provides a three-dimensional spatiotemporal wind field prediction method, system and storage medium for wind turbines based on frequency domain physical information neural network. Background Art

[0002] Wind energy is a renewable, sustainable and clean energy source that has been widely utilized and developed around the world. There is increasing recognition of the importance of wind energy in achieving carbon neutrality goals and meeting global electricity demand (Scholbrock et al., 2016; Murthy and Rahi, 2017; Zhu et al., 2021). However, in wind farms, the randomness of wind fields often leads to intermittent and variability in wind power generation. This variability poses a major challenge to the safe and stable operation of wind turbine structures (Lin et al., 2020; Liu et al., 2023; Cheng et al., 2024). Therefore, accurate and comprehensive wind speed data are essential for evaluating the dynamic response of wind turbine structures and ensuring their operational safety and reliability. Unfortunately, for any specific wind farm or wind turbine structure, the number of actual meteorological observation stations is usually limited, resulting in sparse data distribution in space. This limitation of wind farm data complicates the accurate assessment of wind conditions, thereby posing a major challenge to evaluating the dynamic response of wind turbine structures and ensuring their performance and safety. To address this limitation, conditional simulation of wind farms is essential for reliable data extrapolation to fill existing gaps. Therefore, accurate prediction of three-dimensional spatiotemporal wind fields is essential for evaluating the dynamic response of wind turbine structures and maintaining their safe operation. Enhanced prediction capabilities can help make more informed decisions in the design, layout, and operation strategies of wind turbines, thereby helping to optimize the safety of wind turbine systems.

[0003] To address the challenges associated with spatiotemporal distribution in wind farm forecasting, two main approaches have been proposed to improve the accuracy of spatiotemporal wind forecasting. The first approach uses neural networks to extract and organize spatiotemporal correlations from measurement data and then structure them into datasets to facilitate simultaneous spatiotemporal predictions. For example, Liang et al. (2022) used convolutional neural networks (CNNs) to discern the spatial correlations between measurement points on wind turbines or towers and combined them with multi-layer perceptrons (MLPs) for comprehensive spatiotemporal wind forecasting. Yan et al. (2023) integrated three-dimensional convolutions in a deep learning framework to model the spatiotemporal correlations between wind farm measurement points, achieving remarkable prediction accuracy. Despite the progress made by these methods, they also face inherent challenges. Flattening the spatial information into a one-dimensional vector of the temporal model destroys the intrinsic spatial relationship between the measurement points, which may reduce the spatial correlation. Another approach combines measurement data with principles of fluid dynamics for spatiotemporal wind farm forecasting. Mercieca et al. (2020) derived a reduced-order nonlinear dynamic model by combining the Navier-Stokes equations with the pressure Poisson equation, and then used the unscented Kalman filter (UKF) algorithm to reconstruct the flow field. Zhang et al. (2019, 2021) addressed this problem by embedding a simplified two-dimensional Reynolds-averaged Navier-Stokes (RANS) model in a neural network to reconstruct and predict the two-dimensional wind field without simplification. However, due to the complex dynamic characteristics of real-world wind fields, with strong nonlinear and multi-scale properties, coupled with limited spatial measurement points, the reliance on computational fluid dynamics simulations and simplified Navier-Stokes models limits their applicability. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a three-dimensional space-time wind field prediction method, system and storage medium for wind turbines based on frequency domain physical information neural network, which can accurately predict the three-dimensional space-time wind field, which is crucial for evaluating the dynamic response of the wind turbine structure and ensuring its safe and stable operation.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention first proposes a three-dimensional spatiotemporal wind field prediction method for a wind turbine based on a frequency domain physical information neural network, comprising the following steps:

[0007] Step 1: Collect wind speed data through sensors arranged at multiple measuring points at the wind turbine location;

[0008] Step 2: Preprocess the collected wind speed data, select the wind speed data of some measuring points as the training set, and the wind speed data of the remaining measuring points as the verification set;

[0009] Step 3: Determine the physical model and frequency domain information of the three-dimensional spatiotemporal wind field based on the wind field data of the training set;

[0010] Step 4: Integrate the physical model of the three-dimensional space-time wind field and the frequency domain information into the neural network to construct a frequency domain physical information neural network deep learning model;

[0011] Step 5: Use the training set to train the frequency domain physical information neural network deep learning model; use the validation set to evaluate the generalization ability of the trained frequency domain physical information neural network deep learning model;

[0012] Step 6: The wind speed data is collected in real time by sensors at the measuring points arranged at the wind turbine location. The real-time collected wind speed data is pre-processed and input into the frequency domain physical information neural network deep learning model to predict the three-dimensional spatiotemporal wind field of the wind turbine including the physical model and frequency domain information.

[0013] Furthermore, in step 1, the sampling frequencies and sampling durations of the sensors arranged at all measuring points are equal.

[0014] Furthermore, in step 2, the wind speed data is preprocessed including data missing processing and data outlier processing, and the missing data and outlier data are filled by the mean replacement method, the method is:

[0015]

[0016] Where: x T Indicates the wind speed value at time T; x T-1 represents the wind speed value at time T-1; x T+1 Indicates the wind speed value at time T+1.

[0017] Further, the physical model includes a wind profile model, and the wind profile model is expressed as:

[0018]

[0019] in: Indicates the average wind speed at the reference height; z ref Indicates the reference altitude; represents the average wind speed at height z; α represents the surface roughness index.

[0020] Further, the frequency domain information includes a coherence function model, and the coherence function model adopts the Davenport model, which is expressed as:

[0021]

[0022] Where: γ(f,d jk ) represents the coherence function; S jk(f,d jk ) represents the cross power spectrum of the wind field at measuring point j and measuring point k; S jj (f) and S kk (f) represents the auto-power spectrum of the wind field at measuring point j and measuring point k respectively; d jk represents the Euclidean distance between measuring point j and measuring point k; C y represents the coherence function attenuation factor in the y direction; C z represents the coherence function attenuation factor in the z direction; y j and z j Respectively represent the horizontal and vertical coordinates of the measuring point j; y k and z k Respectively represent the horizontal and vertical coordinates of the measuring point k; Indicates the height z of the measuring point j j The average wind speed at the location; Indicates the height z of the measuring point k k The average wind speed at the location; f represents the frequency; Re[·] represents the real part of the function.

[0023] Furthermore, the frequency domain physical information neural network deep learning model includes a neural network module, a physical information module and a feedback module. The neural network module adopts a fully connected neural network to learn nonlinear relationships; the physical information module adopts the physical model and frequency domain information to train the neural network module; the feedback module adopts a loss function constructed based on the physical model and frequency domain information to reversely transmit the prediction accuracy.

[0024] Furthermore, the loss function is the sum of the supervision loss and the residual of the control equation, expressed as:

[0025] L T =W e L e +W b L b

[0026]

[0027]

[0028] in:

[0029]

[0030]

[0031] Where: L T is the loss function; L e and L b are the loss function components corresponding to the residual of the control equation and the supervision loss of the sampling point respectively; We and W b Respectively represent L e and L b The weight coefficient of e N is the number of points used to calculate the residual of the control equation; b represents the number of training samples in the domain; Refers to the given data of point n in the domain; is the vector predicted by the frequency domain physical information neural network deep learning model at the corresponding point; represents the residual of the ith control equation at point n;

[0032] Indicates the average wind speed at the reference height; z ref Indicates the reference altitude; represents the average wind speed at height z; α represents the surface roughness index;

[0033] Where: S jk (f,d jk ) represents the cross power spectrum of the wind field at measuring point j and measuring point k; S jj (f) and S kk (f) represents the auto-power spectrum of the wind field at measuring point j and measuring point k respectively; d jk represents the Euclidean distance between measuring point j and measuring point k; C y represents the coherence function attenuation factor in the y direction; C z represents the coherence function attenuation factor in the z direction; y j and z j Respectively represent the horizontal and vertical coordinates of the measuring point j; y k and z k Respectively represent the horizontal and vertical coordinates of the measuring point k; Indicates the height z of the measuring point j j The average wind speed at the location; Indicates the height z of the measuring point k k The average wind speed at the location; f represents the frequency; Re[·] represents the real part of the function.

[0034] The present invention also proposes a three-dimensional spatiotemporal wind field prediction system for wind turbines based on frequency domain physical information neural network, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program and implements the three-dimensional spatiotemporal wind field prediction method for wind turbines based on frequency domain physical information neural network as described above.

[0035] The present invention also proposes a storage medium having a computer program stored thereon. When the computer program is executed by a processor, the three-dimensional spatiotemporal wind field prediction method for wind turbines based on frequency domain physical information neural network as described above is implemented.

[0036] The beneficial effects of the present invention are:

[0037] The present invention is a three-dimensional spatiotemporal wind field prediction method for wind turbines based on frequency domain physical information neural network. The frequency domain physical information neural network deep learning model is constructed by integrating the physical model of the three-dimensional spatiotemporal wind field and the frequency domain information into the neural network. The frequency domain physical information neural network deep learning model can establish a robust mapping connection between the physical model of the three-dimensional spatiotemporal wind field and the key wind parameters of the frequency domain information. It is good at predicting the complex frequency domain characteristics of the spatiotemporal wind field and can achieve high-precision prediction with a minimum of measurement points. It can obtain detailed wind field predictions by inputting wind speed data from limited measurement points, and provides a comprehensive tool for environments with limited data collection. The present invention proposes a new wind field prediction method, which improves the prediction capability required for wind energy system optimization and safety, makes decisions on wind turbine design, layout and operation more informed, and helps to effectively and sustainably utilize wind energy. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0039] Figure 1 This is a schematic diagram of the arrangement of wind field prediction measurement points in front of the wind turbine;

[0040] Figure 2 It is a flow chart of an embodiment of a method for predicting a three-dimensional spatiotemporal wind field of a wind turbine based on a frequency domain physical information neural network according to the present invention;

[0041] Figure 3 It is the measured three-dimensional wind field map;

[0042] Figure 4 It is a schematic diagram of the structure of the frequency domain physical information neural network;

[0043] Figure 5 This is the evolution diagram of the loss function during the neural network training process when 50% of the measurement point data is selected as the training set;

[0044] Figure 6 The three-dimensional graph of the FD-PINN wind field prediction results with 50% of the measured data points selected;

[0045] Figure 7 Schematic diagram of the difference between the wind field predicted by FD-PINN and the measured wind field results;

[0046] Figure 8 It is a comparison chart of the predicted value and the true value of the wind speed time history at different heights;

[0047] Fig. 9 Comparison chart of predicted and true values ​​of wind speed autospectrum at different locations. DETAILED DESCRIPTION

[0048] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0049] In this embodiment, in order to evaluate the effectiveness and accuracy of the proposed wind turbine three-dimensional spatiotemporal wind field prediction method based on frequency domain physical information neural network in predicting the three-dimensional spatiotemporal wind field, the method was applied to predict the wind field at the wind turbine location. The longitudinal and transverse spans of the wind field are both 100 meters, with 11 grid points distributed in each direction, and each grid point is a measurement point, that is, there are a total of 121 measurement points for wind speed data on the plane. The schematic diagram of the measurement point layout for wind field prediction in front of the wind turbine is shown in the figure below. Figure 1 shown.

[0050] like Figure 2 As shown, the three-dimensional spatiotemporal wind field prediction method of a wind turbine based on a frequency domain physical information neural network in this embodiment includes the following steps.

[0051] Step 1: Collect wind speed data through sensors at multiple measuring points arranged at the wind turbine location. Specifically, in this embodiment, the sampling frequency and sampling duration of the sensors arranged at all measuring points are equal.

[0052] Step 2: Preprocess the collected wind speed data to obtain Figure 3 The schematic diagram of the measured three-dimensional wind field is shown. The wind speed data of some measuring points are selected as the training set, and the wind speed data of the remaining measuring points are selected as the verification set. In this embodiment, the wind speed data of 50% of the measuring points are selected as the training set, and the wind speed data of the remaining 50% of the measuring points are selected as the verification set.

[0053] In this embodiment, the wind speed data is preprocessed including data missing processing and data outlier processing, and the mean replacement method is used to fill the missing data and outlier data. The method is:

[0054]

[0055] Where: x T Indicates the wind speed value at time T; x T-1 represents the wind speed value at time T-1; x T+1 Indicates the wind speed value at time T+1.

[0056] Step 3: Determine the physical model and frequency domain information of the three-dimensional space-time wind field based on the wind field data in the training set.

[0057] Specifically, the physical model includes a wind profile model, which is expressed as:

[0058]

[0059] in: Indicates the average wind speed at the reference height; z ref Indicates the reference altitude; represents the average wind speed at height z; α represents the surface roughness index.

[0060] The frequency domain information includes a coherence function model, which adopts the Davenport model and is expressed as:

[0061]

[0062] Where: γ(f,d jk ) represents the coherence function; S jk (f,d jk ) represents the cross power spectrum of the wind field at measuring point j and measuring point k; S jj (f) and S kk (f) represents the auto-power spectrum of the wind field at measuring point j and measuring point k respectively; d jk represents the Euclidean distance between measuring point j and measuring point k; C y represents the coherence function attenuation factor in the y direction; C z represents the coherence function attenuation factor in the z direction; y j and z j Respectively represent the horizontal and vertical coordinates of the measuring point j; y k and z k Respectively represent the horizontal and vertical coordinates of the measuring point k; Indicates the height z of the measuring point j j The average wind speed at the location; Indicates the height z of the measuring point k k The average wind speed at the location; f represents the frequency; Re[·] represents the real part of the function.

[0063] Step 4: Integrate the physical model and frequency domain information of the three-dimensional space-time wind field into the neural network to construct a frequency domain physical information neural network deep learning model (FD-PINN).

[0064] In this embodiment, the frequency domain physical information neural network deep learning model includes a neural network module, a physical information module and a feedback module, such as Figure 4As shown. Specifically, in this embodiment, the neural network module adopts a fully connected neural network to learn nonlinear relationships. A fully connected neural network is usually called a multilayer perceptron (MLP), which is used to identify complex functions through a back-propagation algorithm. Specifically, MLP consists of two or more layers of interconnected nodes, including an input layer, multiple hidden layers, and an output layer. Each node in the network is assigned a specific weight and bias. The nodes are fully connected, which means that each node in one layer is linked to each node in the next layer. The input layer is responsible for receiving raw data, while the output layer generates the final result. The hidden layer located between the input layer and the output layer is crucial for feature extraction and performing nonlinear transformations on the input data. A key feature of MLP is its activation function, which enables the network to model and learn nonlinear relationships.

[0065] In this embodiment, the physical information module uses the physical model and frequency domain information to train the neural network module.

[0066] In this embodiment, the feedback module uses a loss function based on the physical model and frequency domain information to reverse the prediction accuracy. Specifically, this embodiment uses spatial and temporal coordinates (x, y, z, t) as input data and u (x, y, z, t) as output parameters. The loss function is the sum of the supervision loss and the residual of the control equation, expressed as:

[0067] L T =W e L e +W b L b

[0068]

[0069]

[0070] in:

[0071]

[0072]

[0073] Where: L T is the loss function; L e and L b are the loss function components corresponding to the residual of the control equation and the supervision loss of the sampling point respectively; W e and W b Respectively represent L e and L b The weight coefficient of e N is the number of points used to calculate the residual of the control equation; b represents the number of training samples in the domain; Refers to the given data of point n in the domain; is the vector predicted by the frequency domain physical information neural network deep learning model at the corresponding point; represents the residual of the ith control equation at point n.

[0074] Step 5: Use the training set to train the frequency domain physical information neural network deep learning model. Use the validation set to evaluate the generalization ability of the trained frequency domain physical information neural network deep learning model. Specifically, input the validation set into the FD-PINN model, evaluate the generalization ability of the trained FD-PINN model, and reversely process the data to obtain the real three-dimensional space-time wind field.

[0075] In this embodiment, the architecture configuration of the FD-PINN model is 5-200-200-200-200-200-1, the hidden layer uses the Tanh activation function, and the output layer uses the linear activation function. The initial learning rate is set to 1×10 -3 In order to accelerate the convergence of the training process, the weight coefficient is set to W e =5 and W b =1. Each training iteration takes about 0.1 seconds. The method proposed in this embodiment achieves real-time 3D wind field reconstruction through offline training and online updating. Then, the model is trained using 50% of the wind field data randomly selected. The deep learning model is trained for 100,000 steps. The evolution of each loss term in the loss function and the total loss is shown in Figure 5 Then, all the spatial coordinate parameters and time of the measurement points are substituted into the trained model to predict the three-dimensional graph of the complete space-time three-dimensional flow field. The prediction results are shown in Figure 6 In order to compare the accuracy of the prediction results of the frequency domain physical information neural network deep learning model, the predicted values ​​are compared with the measured values ​​and the results are shown in Figure 7 As shown. It can be seen that the predicted wind field distribution is very consistent with the actual measurement results, effectively capturing the non-steady-state characteristics of the wind field over time. The error between the two is very small, and the prediction error is usually within 3 meters per second. Only a few cases exceed this threshold, indicating the excellent accuracy of the frequency-domain physical information neural network deep learning model constructed in this embodiment in predicting wind fields, and proving its effectiveness as a reliable tool for capturing the dynamics of wind-induced phenomena. Subsequently, in order to more intuitively compare the difference between the predicted results of the frequency-domain physical information neural network deep learning model and the actual values, wind speed time series at different heights were selected for analysis, and the comparison results are shown in the figure. Figure 8 The results show that the model effectively captures the wind shear phenomenon, i.e., the characteristics of wind speed varying with height. These findings highlight the superior performance of the proposed method and clearly illustrate its effectiveness in accurately predicting wind field characteristics.

[0076] Step 6: The wind speed data is collected in real time by sensors at the measuring points arranged at the wind turbine location. The real-time collected wind speed data is pre-processed and input into the frequency domain physical information neural network deep learning model to predict the three-dimensional spatiotemporal wind field of the wind turbine including the physical model and frequency domain information.

[0077] Fig. 9 The predicted and true values ​​of the wind speed autospectrum at different locations are shown as a function of frequency. It can be seen that the predicted values ​​of the wind speed autospectrum are basically consistent with the true values. Considering that the frequency domain physical information neural network deep learning model takes into account spatial correlation, by comparing the predicted values ​​of the wind speed coherence function with the actual values, the results show that the predicted coherence function is very close to the true value. Even if the distance between the measurement points increases, the model can maintain the overall prediction reliability. This robustness emphasizes the value of the frequency domain physical information neural network deep learning model as a tool for analyzing wind fields at different spatial scales.

[0078] The present embodiment predicts the three-dimensional spatiotemporal wind field of wind turbines based on the frequency domain physical information neural network to predict the three-dimensional spatiotemporal wind field, combines frequency domain information with deep learning technology, and uses wind profile model and wind coherence function model as residual components to improve the prediction accuracy. A notable feature of the method of the present embodiment is that it can achieve high accuracy with a minimum of measurement points, which greatly reduces the reliance on dense and expensive neural networks. The method of the present embodiment is good at predicting the complex frequency domain characteristics of the spatiotemporal wind field by establishing a robust mapping connection between key wind parameters such as wind profile and coherence function, and provides a comprehensive tool for environments with limited data collection by facilitating detailed wind field prediction from limited data input. The method of the present embodiment introduces a new wind field prediction framework, improves the prediction capabilities required for wind energy system optimization and safety, makes decisions about wind turbine design, layout and operation more informed, and contributes to the effective and sustainable use of wind energy. Finally, the method of the present embodiment is applied to the prediction of wind turbine wind field. The longitudinal and lateral spans of the wind field are both 100 meters, 11 grid points are distributed in each direction, and there are a total of 121 wind speed data points on the plane. The research results show that integrating frequency domain information can significantly improve the accuracy of wind turbine spatial wind speed distribution prediction. In addition, the method of this embodiment also reduces the spatial dependence of wind speed. Through verification of real wind fields, the feasibility and accuracy of the FD-PINN model in predicting three-dimensional spatiotemporal wind fields are further confirmed. It should be noted that the three-dimensional spatiotemporal wind field prediction method of wind turbines based on frequency domain physical information neural network in this embodiment is not limited to wind turbine structures, but also for other structures such as photovoltaic structures, tall buildings, large-span cable structures and large-span bridges.

[0079] This embodiment also proposes a three-dimensional spatiotemporal wind field prediction system for wind turbines based on frequency domain physical information neural network, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program and implements the three-dimensional spatiotemporal wind field prediction method for wind turbines based on frequency domain physical information neural network as described above in this embodiment.

[0080] The present invention also proposes a storage medium having a computer program stored thereon. When the computer program is executed by a processor, the three-dimensional spatiotemporal wind field prediction method for a wind turbine based on a frequency domain physical information neural network as described above in this embodiment is implemented.

[0081] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for predicting the three-dimensional spatiotemporal wind field of a wind turbine based on a frequency domain physical information neural network, characterized in that: The steps include: Step 1: Collect wind speed data through sensors arranged at multiple measuring points at the wind turbine location; Step 2: Preprocess the collected wind speed data, select the wind speed data of some measuring points as the training set, and the wind speed data of the remaining measuring points as the verification set; Step 3: Determine the physical model and frequency domain information of the three-dimensional spatiotemporal wind field based on the wind field data of the training set; Step 4: Integrate the physical model of the three-dimensional space-time wind field and the frequency domain information into the neural network to construct a frequency domain physical information neural network deep learning model; Step 5: Use the training set to train the frequency domain physical information neural network deep learning model; use the validation set to evaluate the generalization ability of the trained frequency domain physical information neural network deep learning model; Step 6: The wind speed data is collected in real time by sensors at the measuring points arranged at the wind turbine location. The real-time collected wind speed data is pre-processed and input into the frequency domain physical information neural network deep learning model to predict the three-dimensional spatiotemporal wind field of the wind turbine including the physical model and frequency domain information.

2. The method for predicting the three-dimensional spatiotemporal wind field of a wind turbine based on a frequency domain physical information neural network according to claim 1 is characterized in that: In the step 1, the sampling frequencies and sampling durations of the sensors arranged at all measuring points are equal.

3. The method for predicting the three-dimensional spatiotemporal wind field of a wind turbine based on a frequency domain physical information neural network according to claim 1 is characterized in that: In the step 2, the wind speed data is preprocessed including data missing processing and data outlier processing, and the mean replacement method is used to fill the missing data and outlier data, the method is: Where: x T Indicates the wind speed value at time T; x T―1 represents the wind speed value at time T-1; x T+1 Indicates the wind speed value at time T+1.

4. The method for predicting a wind turbine's three-dimensional spatiotemporal wind field based on a frequency domain physical information neural network according to claim 1 is characterized in that: The physical model includes a wind profile model, which is expressed as: in: Indicates the average wind speed at the reference height; z ref Indicates the reference altitude; represents the average wind speed at height z; α represents the surface roughness index.

5. The method for predicting the three-dimensional spatiotemporal wind field of a wind turbine based on a frequency domain physical information neural network according to claim 1 is characterized in that: The frequency domain information includes a coherence function model, and the coherence function model adopts the Davenport model, which is expressed as: Where: γ(f,d jk ) represents the coherence function; S jk (f,d jk ) represents the cross power spectrum of the wind field at measuring point j and measuring point k; S jj (f) and S kk (f) represents the auto-power spectrum of the wind field at measuring point j and measuring point k respectively; d jk represents the Euclidean distance between measuring point j and measuring point k; C y represents the coherence function attenuation factor in the y direction; C z represents the coherence function attenuation factor in the z direction; y j and z j Respectively represent the horizontal and vertical coordinates of the measuring point j; y k and z k Respectively represent the horizontal and vertical coordinates of the measuring point k; Indicates the height z of the measuring point j j The average wind speed at the location; Indicates the height z of the measuring point k k The average wind speed at the location; f represents the frequency; Re[·] represents the real part of the function.

6. The method for predicting the three-dimensional spatiotemporal wind field of a wind turbine based on a frequency domain physical information neural network according to claim 1, characterized in that: The frequency domain physical information neural network deep learning model includes a neural network module, a physical information module and a feedback module. The neural network module uses a fully connected neural network to learn nonlinear relationships. The physical information module uses the physical model and frequency domain information to train the neural network module. The feedback module uses a loss function constructed based on the physical model and frequency domain information to reversely transmit the prediction accuracy.

7. The method for predicting the three-dimensional spatiotemporal wind field of a wind turbine based on a frequency domain physical information neural network according to claim 6, characterized in that: The loss function is the sum of the supervision loss and the residual of the control equation, expressed as: L T =W e L e +W b L b in: Where: L T is the loss function; L e and L b are the loss function components corresponding to the residual of the control equation and the supervision loss of the sampling point respectively; W e and W b Respectively represent L e and L b The weight coefficient of e N is the number of points used to calculate the residual of the control equation; b represents the number of training samples in the domain; Refers to the given data of point n in the domain; is the vector predicted by the frequency domain physical information neural network deep learning model at the corresponding point; represents the residual of the ith control equation at point n; Indicates the average wind speed at the reference height; z ref Indicates the reference altitude; represents the average wind speed at height z; α represents the surface roughness index; Where: S jk (f,d jk ) represents the cross power spectrum of the wind field at measuring point j and measuring point k; S jj (f) and S kk (f) represents the auto-power spectrum of the wind field at measuring point j and measuring point k respectively; d jk represents the Euclidean distance between measuring point j and measuring point k; C y represents the coherence function attenuation factor in the y direction; C z represents the coherence function attenuation factor in the z direction; y j and z j Respectively represent the horizontal and vertical coordinates of the measuring point j; y k and z k Respectively represent the horizontal and vertical coordinates of the measuring point k; Indicates the height z of the measuring point j j The average wind speed at the location; Indicates the height z of the measuring point k k The average wind speed at the location; f represents the frequency; Re[·] represents the real part of the function.

8. A wind turbine three-dimensional spatiotemporal wind field prediction system based on frequency domain physical information neural network, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program and implements the three-dimensional spatiotemporal wind field prediction method for a wind turbine based on a frequency domain physical information neural network as described in any one of claims 1 to 7.

9. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements a three-dimensional spatiotemporal wind field prediction method for a wind turbine based on a frequency domain physical information neural network as described in any one of claims 1 to 7.

Images