Mountain area sudden change wind field active control wind tunnel simulation method based on deep learning

Abstract

The invention discloses a mountainous area sudden change wind field active control wind tunnel simulation method based on deep learning, and relates to the technical field of wind tunnel experiment research, and the method comprises the steps: obtaining a target wind field parameter in a mountainous area target region, constructing an MLP input vector according to the target wind field parameter in combination with wind tunnel real-time sensor data, a deep learning control model comprising an MLP forward mapping model and a Transform time sequence feedback model is established, the MLP forward mapping model generates a multi-fan array control signal according to the MLP input vector, a multi-fan array is controlled to generate a simulated wind field, the Transform time sequence feedback model generates a multi-fan array control correction signal according to a wind field difference value between the simulated wind field and a target wind field, and the multi-fan array control correction signal is controlled to generate a multi-fan array control correction signal according to a wind field difference value between the simulated wind field and the target wind field. The simulated wind field parameters are dynamically corrected in real time, and target wind field parameters are approached; according to the method, high-fidelity dynamic reproduction and intelligent adaptive control of the complex wind field in the mountainous area are realized through MLP feedforward generation and Transform feedback correction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind tunnel experimental research, in particular to a mountain sudden wind field active control wind tunnel simulation method based on deep learning. BACKGROUND

[0002] The wind field under the condition of complex terrain in mountainous areas has significant time non-stationarity, spatial non-uniformity and strong suddenness characteristics. Mountain wind field often shows complex flow structure of gust, wind shear and fluctuating turbulence, which poses a major challenge to the safety and service life of engineering facilities such as bridge structures, power transmission lines and wind turbines. Therefore, how to truly reproduce the characteristics of sudden wind field in mountainous areas in wind tunnel test and provide a basis for studying and verifying the response behavior of engineering structures under sudden wind field in mountainous areas has become an important research direction in the field of wind engineering and structural wind resistance.

[0003] Traditional wind tunnel simulation methods mostly use passive grids, rotating blades or spoilers to generate turbulent wind field, but such methods have the problems of poor controllability of wind field, limited wind field reconstruction ability and inability to intelligently respond to changes in wind field. In the prior art, active control wind tunnel has become an important means to simulate non-uniform wind field. A typical scheme is to use a multi-fan array arrangement to generate different wind speed distribution and turbulence characteristics by using an independently controlled fan group. For example, patent document CN108254151A discloses a multi-fan active control tornado wind tunnel, which simulates tornado vortex by combined operation of multiple fans; patent document CN119935481A proposes an active control wind tunnel that simulates average wind and fluctuating wind separately, which can achieve component superposition of wind field to a certain extent. However, the above methods are mainly aimed at specific types of wind field and are difficult to cope with sudden wind field, turbulence or wind shear characteristics under complex terrain in mountainous areas, and cannot meet the demand for dynamic and real-time reproduction of complex wind field in mountainous areas. SUMMARY

[0004] The purpose of the present application is to provide a mountain sudden wind field active control wind tunnel simulation method based on deep learning to solve the problem that the traditional wind tunnel simulation method proposed in the background art is difficult to dynamically and real-time reproduce the characteristics of complex terrain wind field in mountainous areas.

[0005] To achieve the above purpose, the present application provides the following technical scheme: A mountain sudden wind field active control wind tunnel simulation method based on deep learning, comprising the following steps: S1: deploying an anemometer array in a target mountain area to obtain target wind field parameters of the target mountain area, performing mutation identification on the target wind field parameters to construct a mutation feature vector; constructing a wind field feature vector according to the target wind field parameters; performing principal component analysis (PCA) and mutual information screening on the wind field feature vector and wind tunnel real-time sensor data to construct an optimal feature subset, combining the mutation feature vector and the optimal feature subset to construct an MLP input vector; S2: establishing a deep learning control model, including a forward mapping model and a feedback optimization model, the forward mapping model adopts a multi-layer perceptron (MLP), and the feedback optimization model adopts a time-series feedback model based on a Transformer neural network; the MLP is used to establish a mapping relationship between the target wind field parameters and control signals of a multi-fan array in a wind tunnel, and the Transformer neural network is used to dynamically correct deviations between simulated wind field parameters in a wind tunnel test section and the target wind field parameters; S3: inputting the MLP input vector into the MLP to generate multi-fan array control signals; S4: a multi-fan array control system operates a multi-fan array according to the multi-fan array control signals to generate a simulated wind field in a wind tunnel test section; S5: a wind speed and direction sensor array is used to collect wind field real-time data of the simulated wind field, and a wind field difference between the wind field real-time data and the target wind field parameters is calculated; S6: the Transformer neural network establishes an error evolution law model according to the wind field difference and historical control data, predicts an error evolution trend, and generates a multi-fan array control correction signal according to the error evolution trend; S7: the multi-fan array control system adjusts the rotation speed and inclination angle of the multi-fan array according to the multi-fan array control correction signal; S8: repeating steps S5 to S7 until the wind field difference meets a preset stability criterion.

[0006] The principle of this invention, a deep learning-based active control wind tunnel simulation method for abrupt wind fields in mountainous areas, is as follows: An array of anemometers and wind vanes is deployed in the target area of ​​the mountainous region to collect target wind field parameters. A wind field feature vector is constructed through feature recognition, principal component analysis, and mutual information filtering. A multilayer perceptron (MLP) is used to establish a mapping between the wind field feature vector and multi-fan control signals. Based on the target wind field feature vector, a multi-fan array control signal is generated to control the multi-fan array to generate a simulated wind field in the wind tunnel test section. Simultaneously, a feedback optimization model based on a Transformer neural network is introduced to predict the wind field error evolution trend and generate a multi-fan array control correction signal to dynamically correct the deviation between the simulated wind field and the target wind field, thereby achieving high-precision dynamic simulation of abrupt wind fields in mountainous areas. This method employs a deep learning control model combining forward mapping and temporal feedback optimization to generate and dynamically correct the simulated wind field based on the target wind field parameters. This not only improves the response speed and control accuracy of wind tunnel simulation but also solves the problems of poor responsiveness and stability in traditional wind tunnel simulation methods when dealing with abrupt wind fields, thus improving the realism and stability of the simulated wind field.

[0007] Preferably, to comprehensively and efficiently obtain the real target wind field parameters of complex mountainous terrain, in step S1, the deployment method of the anemometer array is as follows: the target area of ​​the mountainous region is divided into an unencrypted area and an encrypted area according to CFD simulation, and the array is deployed in the unencrypted area according to a preset density. Deploy anemometers of wind speed and direction in the densely populated area according to a preset deployment density. Deploy anemometers and wind direction indicators. < To improve the efficiency of wind field parameter acquisition and the spatial resolution of wind field parameters in key airflow areas; the densification zone includes: airflow acceleration zone, separation zone and reattachment zone.

[0008] Preferably, in order to reduce the impact of environmental and instrument noise on the MLP input vector and to achieve abrupt feature extraction of the target wind field parameters, the abrupt feature identification operation in step S1 includes the following steps: S11: Kalman filtering is used to denoise the target wind field parameters to obtain wind field preprocessing data; S12: Employing a mutation identification algorithm based on wavelet packet energy entropy or variational mode decomposition (VMD), the wind speed sequence in the wind field preprocessing data is decomposed into multiple frequency bands according to a preset frequency band range and number of frequency bands. This detects gust mutation intervals and energy distribution changes, extracts mutation features to form the mutation feature vector, which includes: wind speed amplitude mutation features, wind speed temporal mutation features, wind speed energy distribution features, turbulence features, and spatial correlation features.

[0009] Preferably, to achieve the adaptive capability of the mutation identification algorithm to cope with different wind field distribution changes and improve the accuracy and generalization ability of feature extraction, the mutation identification operation introduces an online self-learning mechanism. Based on the distribution changes of the target wind field parameters, it automatically adjusts the denoising parameters of the Kalman filter, the preset frequency band range and number of frequency bands, and the parameter composition of the mutation feature vector. The wind speed amplitude mutation features include: the wind speed value at the start of the mutation, the peak wind speed of the mutation, the rate of change of wind speed amplitude, and the wind speed value at the end of the mutation. The wind speed temporal mutation features include: the duration of gust mutations. The wind field feature vector includes: the delay time between the peak wind speed occurrence and the start time of the abrupt change; the duration of the wind speed rise phase; and the duration of the wind speed fall phase. The wind speed energy distribution characteristics include: the proportion of high-frequency energy and the extreme values ​​of wavelet packet energy entropy within the abrupt change interval. The turbulence characteristics include: the turbulence intensity within the abrupt change interval; the peak value of the fluctuating wind speed power spectrum in the abrupt change frequency band; and the maximum change in the wind angle of attack. The spatial correlation characteristics include: the wind speed difference between adjacent measuring points within the abrupt change interval and the wind direction deviation. The parameters of the wind field feature vector include: the wind speed component along the horizontal mainstream direction of the target mountain area. Horizontal lateral wind speed component perpendicular to the horizontal mainstream direction Vertical wind speed component perpendicular to the horizontal mainstream direction Parameters including wind angle of attack, wind shear, turbulence intensity, and power spectrum of fluctuating wind speed.

[0010] Preferably, to achieve collaborative training of the Multilayer Perceptron (MLP) and the Transformer Neural Network (NNN) model, enabling the deep learning control model to simultaneously optimize initial control and error correction capabilities during the training phase, thereby improving the control accuracy of wind speed and direction for the multi-fan array, after each round of wind tunnel experiments, the deep learning control model uses a joint loss function to collaboratively train the MLP and the Transformer NN. The control signal output by the MLP serves as the initial input state for the Transformer NN, which corrects its weight parameters through time-series feedback. The MLP employs a loss function that includes a wind tunnel wind speed distribution error term during the training phase. The joint loss function is expressed as follows: ; ; ; In the formula, For the joint loss function, and These are the loss terms of the Multilayer Perceptron (MLP) and the Transformer Neural Network, respectively. and These are the weight coefficients for the MLP loss term and the Transformer loss term, respectively. These are the control commands output by the multilayer perceptron (MLP). For target control commands, The wind speed distribution is formed after the fan array is driven by the control commands output by the multilayer sensor (MLP). For the target wind speed distribution, The wind field error predicted by the Transformer neural network. This represents the actual wind field error. These are the control commands modified by the Transformer neural network.

[0011] Preferably, to improve the physical consistency of the simulated wind field, the Transformer neural network introduces wind tunnel flow field conservation constraints and energy constraints into the attention weight calculation formula to ensure that the generated wind speed vector field satisfies the divergence condition and energy conservation condition. The attention weight calculation formula is expressed as follows: ; ; ; ; In the formula, Here is the formula for calculating the attention weights in a Transformer neural network. , and These are the query vector, key vector, and value vector, respectively. For normalization function, For transpose operation, Let be the dimension of the key vector. and These are the weighting coefficients for the flow field conservation constraint term and the energy constraint term, respectively. and These are respectively the flow field conservation constraint and the energy constraint. This represents the wind speed vector field in the wind tunnel. Let be the divergence of the wind speed vector field in the wind tunnel. The total kinetic energy of the wind field in the wind tunnel. The total kinetic energy of the target wind field air density, This is the space area for the wind tunnel test section. This represents the wind speed component along the horizontal extension direction of the wind tunnel. The horizontal lateral wind speed component is perpendicular to the horizontal extension direction of the wind tunnel. The vertical wind speed component is perpendicular to the horizontal extension direction of the wind tunnel.

[0012] Preferably, to enhance the generalization ability of the deep learning control model for wind field parameters of different terrains and to quickly generate initial control strategies applicable to different mountain wind field types, thereby improving the response efficiency of the wind tunnel control system, the deep learning control model introduces a control strategy knowledge distillation mechanism during the training of the Transformer neural network. This mechanism extracts hidden features from the historical control data and transfers these features to the encoding layer of the current Transformer neural network. The historical control data includes control command sequences and wind field response information recorded during all wind tunnel experiments prior to the current wind tunnel experiment. The hidden features serve as auxiliary information during the training process and participate in the parameter updates of the current Transformer neural network. The deep learning control model pre-configures control strategy templates for multiple typical mountain wind field types, including canyon gusts, mountaintop wind shear, and plateau pulse turbulence. These templates are used to control the multi-fan array to generate the initial wind field in the wind tunnel test section.

[0013] Preferably, to achieve spatial zoning control of the simulated wind field and to respond quickly and accurately to wind speed and direction correction signals, the multi-fan array includes multiple independently controllable fans arranged upstream of the wind tunnel test section. These fans are arranged in a preset two-dimensional grid array. The fans are divided into a core area array fan group, a core area peripheral array fan group, and an edge area array fan group according to a preset ratio. The core area array fan group is used to simulate sudden changes in the wind field, the core area peripheral array fan group is used for background flow field compensation, and the edge area array fan group is used to provide basic wind speed support. Each of the multiple independently controllable fans is equipped with an independent servo drive device, which independently adjusts the speed and tilt angle of each fan according to the initial multi-fan array control signal or the corrected multi-fan array control signal.

[0014] Preferably, in order to achieve stable operation of the simulated wind field and suppress wind field errors, the multi-fan array control system includes an adaptive threshold judgment mechanism. The adaptive threshold judgment mechanism is used to switch the operation mode of the multi-fan array control system according to the changing trends of wind speed amplitude error, wind direction deviation and turbulence intensity error. The operation mode includes a sudden change response mode and a steady state maintenance mode.

[0015] Preferably, to accurately determine the stability of the simulated wind field, the wind field difference includes: wind speed amplitude error, wind direction deviation, and turbulence intensity error; the preset stability criteria include: the wind speed amplitude error is less than a preset wind speed amplitude error threshold, the wind direction deviation is less than a preset wind direction deviation threshold, and the turbulence intensity error is less than a preset turbulence intensity error threshold.

[0016] One or more technical solutions provided by this invention have at least the following technical effects or advantages: 1. By constructing a deep learning control model that includes a multilayer perceptron (MLP) forward mapping model and a Transformer neural network time-series feedback model, and combining the acquisition of target wind field parameters and feature extraction in mountainous areas to generate simulated wind fields, the wind field error is predicted by the Transformer time-series feedback model and the simulated wind field parameters are iteratively corrected. This enables accurate simulation and dynamic real-time adjustment of characteristic wind fields in mountainous areas in a wind tunnel, thereby improving the stability and accuracy of the simulated wind fields. 2. CFD simulation was used to divide the encrypted and unencrypted areas, and anemometers with differentiated deployment density were deployed to ensure the spatial resolution of wind field parameters in key airflow areas, while taking into account the efficiency and effectiveness of parameter acquisition. 3. By using Kalman filtering for noise reduction and wavelet packet energy entropy or VMD algorithm to identify abrupt change intervals, multi-band identification and multi-dimensional feature extraction of wind speed abrupt change intervals and capacity distribution wind are achieved, providing a reliable abrupt change information input vector for MLP; 4. By dynamically adjusting the denoising parameters and abrupt change feature extraction parameters through an online self-learning mechanism, the adaptability and accuracy of abrupt change feature identification to changes in target wind field parameters are improved; 5. By coupling the MLP forward mapping with the Transformer feedback correction process through a joint loss function, the control model can simultaneously optimize the initial control capability and error correction capability during the training phase. Furthermore, by incorporating the wind speed distribution error term into the MLP loss function, the accuracy of the MLP output control signal and the matching degree of the Transformer feedback correction are effectively improved, thereby enhancing the overall control performance of the deep learning control model. 6. By introducing flow field conservation constraints and energy constraints into the attention weight calculation of Transformer, we ensure that the correction signal generated by Transformer conforms to the physical laws of wind field, thereby improving the physical consistency of simulated wind field and the accuracy of Transformer feedback correction. 7. By transferring the hidden features of historical control data to the current Transformer model through the knowledge distillation mechanism, and combining them with pre-configured typical mountain wind field control strategy templates, the model is made capable of cross-scenario adaptation and fast convergence, thereby improving the model's adaptability and response efficiency to different mountain wind fields. 8. Through the functional zoning design of the multi-fan array and independent servo drive, the wind field parameters in different areas can be precisely controlled, thereby improving the spatial distribution accuracy and dynamic response capability of the simulated wind field. 9. By using an adaptive threshold judgment mechanism to dynamically switch between sudden response and steady-state maintenance based on the trend of wind field parameter error changes, the control strategy is ensured to guarantee both the rapid response of the system when sudden wind field occurs and the operational stability under steady-state wind field, thereby improving the operating efficiency of the multi-fan array control system. 10. By clarifying the specific indicators of wind field difference and the quantitative standards of preset stability criteria, we can provide a clear and executable basis for the stability assessment of simulated wind fields, ensure that the wind tunnel simulation results meet the preset accuracy requirements, and guarantee the accuracy and efficiency of wind field simulation. Attached Figure Description

[0017] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention. Figure 1 This is the overall framework diagram of the wind tunnel system for active control of sudden wind field changes in mountainous areas according to the present invention; Figure 2 This is a schematic diagram of the wind tunnel simulation method for active control of sudden wind field in mountainous areas based on deep learning, as proposed in this invention. Figure 3 Schematic diagram of multi-fan array arrangement; Figure 4 This is a schematic diagram of the operation and control of the wind tunnel test section; Figure 5 This is a schematic diagram of a multilayer perceptron (MLP) structure. Figure 6 This is a schematic diagram of the Transformer neural network structure; Among them, 1-multi-fan array, 2-damping net, 3-wind tunnel test section, 4-high-speed wind speed and direction sensor, 5-wind field error calculation system, 6-Transformer neural network time-series feedback model, 7-multi-fan array control computer, 8-PXI system, 9-AC servo, 11-core area array fan group, 12-core area peripheral array fan group, 13-edge area array fan group. Detailed Implementation

[0018] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0020] Example 1 Please refer to Figures 1-6 This invention provides a deep learning-based method for actively controlling wind tunnel simulation of abrupt wind fields in mountainous areas, comprising the following steps: S1: Deploy an array of anemometers in the mountainous target area to obtain the target wind field parameters of the mountainous target area, perform abrupt change identification on the target wind field parameters, and construct abrupt change feature vector; construct a wind field feature vector based on the target wind field parameters; perform principal component analysis (PCA) and mutual information filtering on the wind field feature vector and real-time wind tunnel sensor data to construct an optimal feature subset, and combine the abrupt change feature vector with the optimal feature subset to construct an MLP input vector; S2: Establish a deep learning control model, including a forward mapping model and a feedback optimization model. The forward mapping model adopts a multilayer perceptron (MLP), and the feedback optimization model adopts a time-series feedback model based on a Transformer neural network. The MLP is used to establish the mapping relationship between the target wind field parameters and the control signals of the multi-fan array in the wind tunnel. The Transformer neural network is used to dynamically correct the deviation between the simulated wind field parameters and the target wind field parameters in the wind tunnel test section. S3: Input the MLP input vector into the multilayer perceptron MLP to generate a multi-fan array control signal; S4: The multi-fan array control system operates the multi-fan array according to the multi-fan array control signal to generate a simulated wind field in the wind tunnel test section; S5: Collect real-time wind field data of the simulated wind field through an array of wind speed and direction sensors, and calculate the wind field difference between the real-time wind field data and the target wind field parameters; S6: The Transformer neural network establishes an error evolution law model based on the wind field difference and historical control data, predicts the error evolution trend, and generates a multi-fan array control correction signal based on the error evolution trend; S7: The multi-fan array control system adjusts the speed and tilt angle of the multi-fan array according to the multi-fan array control correction signal; S8: Repeat steps S5 to S7 until the wind field difference meets the preset stability criterion.

[0021] The deployment method of the anemometer array is as follows: based on CFD simulation, the mountainous target area is divided into an unencrypted zone and an encrypted zone. The encrypted zone includes an airflow acceleration zone, a separation zone, and a reattachment zone. The unencrypted zone includes the area of ​​the mountainous target area other than the encrypted zone. Within the unencrypted zone, a preset deployment density is used. Deploy anemometers of wind speed and direction in the densely populated area according to a preset deployment density. Deploy anemometers and wind direction indicators. < ;in, and The value of varies depending on the type of mountainous area and the characteristics of the wind field. Taking the gust wind field of a typical canyon terrain as an example... , In this invention, three layers of measuring points are arranged vertically in the encrypted area, with a vertical distance of 5m between each layer of measuring points, to form a three-dimensional monitoring and acquisition network, so as to achieve precise capture of the three-dimensional structure of complex mountain wind fields. Each measuring point is equipped with a wind speed and direction instrument. CFD simulation is an existing technology in this field. Those skilled in the art can choose the appropriate Fluent software to perform CFD simulation. This invention will not elaborate further.

[0022] The mutation identification operation includes the following steps: S11: Kalman filtering is used to denoise the target wind field parameters to obtain wind field preprocessing data; S12: Employing a mutation identification algorithm based on wavelet packet energy entropy or variational mode decomposition (VMD), the wind speed sequence in the wind field preprocessing data is decomposed into multiple frequency bands according to a preset frequency band range and number of frequency bands. This detects gust mutation intervals and energy distribution changes, extracts mutation features to form the mutation feature vector, which includes: wind speed amplitude mutation features, wind speed temporal mutation features, wind speed energy distribution features, turbulence features, and spatial correlation features. Among them, the process noise covariance of the preset Kalman filter Observation noise covariance Kalman filtering is an existing technology and will not be described in detail in this invention. The mutation identification algorithm based on wavelet packet energy entropy is adopted, the db4 wavelet basis function is selected, the preset frequency band range is 0.1-10Hz, the number of frequency bands is set to 8, and the preset frequency band range is divided into 8 frequency band intervals. The db4 wavelet basis function is an existing technology and will not be described in detail in this invention.

[0023] The mutation identification operation incorporates an online self-learning mechanism, automatically adjusting the denoising parameters of the Kalman filter, the preset frequency band range and number of frequency bands, and the parameter composition of the mutation feature vector based on the distribution changes of the target wind field parameters. Specifically, when the standard deviation of the target wind field parameters exceeds 5% for 5 consecutive minutes, the system automatically adjusts the observation noise covariance of the Kalman filter. The wavelet packet frequency band range is extended to 0.1-12Hz; The wind speed amplitude abrupt change characteristics include: wind speed value at the start of the abrupt change, peak wind speed, rate of change of wind speed amplitude, and wind speed value at the end of the abrupt change; the wind speed temporal abrupt change characteristics include: duration of gust abrupt change, delay time of peak wind speed occurrence relative to the start of the abrupt change, duration of wind speed rise phase, and duration of wind speed fall phase; the wind speed energy distribution characteristics include: high-frequency energy proportion and extreme values ​​of wavelet packet energy entropy within the abrupt change interval; the turbulence characteristics include: turbulence intensity within the abrupt change interval, peak value of fluctuating wind speed power spectrum in the abrupt change frequency band, and maximum change in wind angle of attack; the spatial correlation characteristics include: wind speed difference and wind direction deviation between adjacent measuring points within the abrupt change interval; the parameters of the wind field feature vector include: wind speed component along the horizontal mainstream direction of the target area in the mountainous region. Horizontal lateral wind speed component perpendicular to the horizontal mainstream direction Vertical wind speed component perpendicular to the horizontal mainstream direction Parameters including wind angle of attack, wind shear, turbulence intensity, and power spectrum of fluctuating wind speed.

[0024] In this process, after each round of wind tunnel experiments, the deep learning control model uses a joint loss function to collaboratively train the Multilayer Perceptron (MLP) and the Transformer Neural Network (TN). The control signal output by the MLP serves as the initial input state for the TN, and the TN corrects its weight parameters through time-series feedback. The MLP employs a loss function that includes a wind tunnel wind speed distribution error term during the training phase. The joint loss function is expressed as follows: ; ; ; middle, For the joint loss function, and These are the loss terms of the Multilayer Perceptron (MLP) and the Transformer Neural Network, respectively. and These are the weight coefficients for the MLP loss term and the Transformer loss term, respectively. , , These are the control commands output by the multilayer perceptron (MLP). The target control command is determined based on wind tunnel calibration data. This item is used to measure the difference between the control commands output by the MLP and the target control commands. The wind speed distribution is formed after the fan array is driven by the control commands output by the multilayer sensor (MLP). For the target wind speed distribution, This term measures the difference between the spatial distribution of wind speed generated after the fan array is driven by control commands and the target spatial distribution of wind speed. The wind field error predicted by the Transformer neural network. This represents the actual wind field error. This term measures the difference between the wind field error predicted by the Transformer neural network and the actual wind field error. The control commands are modified by the Transformer neural network. The term is used to measure the difference between the control command corrected by the Transformer neural network and the target control command.

[0025] The Transformer neural network incorporates wind tunnel flow field conservation and energy constraints into its attention weight calculation formula, which is expressed as follows: ; ; ; ; In the formula, Here is the formula for calculating the attention weights in a Transformer neural network. , and These are the query vector, key vector, and value vector, respectively. For normalization function, For transpose operation, Let be the dimension of the key vector. and These are the weighting coefficients for the flow field conservation constraint term and the energy constraint term, respectively. , , and These are respectively the flow field conservation constraint and the energy constraint. This represents the wind speed vector field in the wind tunnel. Let be the divergence of the wind speed vector field in the wind tunnel. The divergence value used to constrain the spatial distribution of wind speed. The total kinetic energy of the wind field in the wind tunnel. The total kinetic energy of the target wind field Used to constrain the difference between the total energy of the simulated wind field and the total energy of the target wind field. air density, This is the space area for the wind tunnel test section. This represents the wind speed component along the horizontal extension direction of the wind tunnel. The horizontal lateral wind speed component is perpendicular to the horizontal extension direction of the wind tunnel. The vertical wind speed component is perpendicular to the horizontal extension direction of the wind tunnel.

[0026] The deep learning control model incorporates a control policy knowledge distillation mechanism during the Transformer neural network training process. This mechanism extracts hidden features from historical control data using a ResNet50 network and transfers these features to the encoding layer of the current Transformer neural network to accelerate training convergence. The historical control data includes control command sequences and wind field response information recorded during all wind tunnel experiments prior to the current experiment. These hidden features serve as auxiliary information during training and participate in parameter updates of the current Transformer neural network. The deep learning control model pre-configures control policy templates for various typical mountain wind field types, including canyon gusts, etc. The control strategy templates for various typical mountain wind field types, including mountaintop wind shear and plateau pulse turbulence, are used to control the generation of the initial wind field in the wind tunnel test section by the multi-fan array. Specifically, the control strategy template for canyon gusts is: initial fan speed of 1500 rpm and initial tilt angle of 3° for the core area fan group; initial fan speed of 1200 rpm and initial tilt angle of 1° for the outer core area fan group; and initial fan speed of 1000 rpm and initial tilt angle of 0° for the edge area fan group. The control strategy template for mountaintop wind shear is: initial fan speed of 1400 rpm and initial tilt angle ±5° for the core area fan group; and initial fan speed of 1100 rpm and initial tilt angle ±2° for the outer core area fan group. The control strategy template for plateau pulse turbulence is: initial fan speed of 1600 rpm and initial tilt angle ±3° for the core area fan group; and initial fan speed of 1300 rpm and initial tilt angle ±1° for the outer core area fan group.

[0027] The multi-fan array includes multiple independently controllable fans arranged upstream of the wind tunnel test section. These fans are arranged in a preset two-dimensional grid array. The minimum number of independently controllable fans in the multi-fan array is 16. In this embodiment, 32 independently controllable fans are used to form a 4×8 vertical plane rectangular grid multi-fan array. Figure 3 As shown; the multiple independently controllable fans are divided into a core area array fan group, a core area peripheral array fan group, and an edge area array fan group according to a preset ratio. In this embodiment, the core area array fan group (corresponding to the key measurement area of ​​the test section) is equipped with 16 high-power, high-dynamic-response fans, each equipped with a high-torque servo motor and carbon fiber composite blades to ensure that the fan array responds quickly and accurately to adjustments in wind speed and direction. The core area peripheral array fan group is equipped with 10 medium-power fans, and the edge area array fan group is equipped with 6 high-power fans. The core area array fan group is used to simulate sudden changes in the wind field, the core area peripheral array fan group is used for background flow field compensation, and the edge area array fan group is used to provide basic wind speed support. Each of the multiple independently controllable fans is equipped with an independent servo drive device. The independent servo drive device is used to independently adjust the speed and tilt angle of each fan according to the initial multi-fan array control signal or the corrected multi-fan array control signal. In this embodiment, the independent servo drive device uses a Panasonic A6 series AC servo motor as the control motor for each fan.

[0028] The multi-fan array control system includes an adaptive threshold judgment mechanism. This mechanism switches the operating mode of the multi-fan array control system based on the changing trends of wind speed amplitude error, wind direction deviation, and turbulence intensity error. The operating modes include a sudden change response mode and a steady-state holding mode. The trigger condition for the sudden change response mode is: wind speed amplitude error... Or wind direction deviation The trigger condition for the steady-state maintenance mode is: wind speed amplitude error. And wind direction deviation And turbulence intensity error In the sudden change response mode, the fan speed adjustment range increases, with a single speed adjustment of ±50 rpm; in the steady state holding mode, the fan speed adjustment range decreases, with a single speed adjustment of ±5 rpm.

[0029] The wind field difference includes: wind speed amplitude error, wind direction deviation, and turbulence intensity error; the preset stability criteria include: the wind speed amplitude error is less than a preset wind speed amplitude error threshold, the wind direction deviation is less than a preset wind direction deviation threshold, and the turbulence intensity error is less than a preset turbulence intensity error threshold; wherein the preset wind speed amplitude error threshold is 1%, the preset wind direction deviation threshold is 0.3°, and the preset turbulence intensity error threshold is 5%.

[0030] Example 2 Based on Example 1, Example 2 will be described and illustrated with specific implementation cases.

[0031] Implementation Case 1: Simulation of Sudden Gusts in Bridges and Canyons The wind resistance performance of a high-altitude canyon bridge under strong gusts needs to be verified in a wind tunnel. This invention's system first collects long-term measured wind speed data from the canyon where the bridge is located, including several typical sudden gust events. The measured wind speed data is then processed using Kalman filtering for noise reduction and wavelet packet energy entropy-based abrupt change feature extraction. This data is then input into a trained MLP forward mapping model to generate a multi-fan array control signal. After the experiment begins, the fan array operates according to the multi-fan array control signal, creating a simulated wind field similar to strong gusts in an actual canyon within the test section. Wind speed sensors distributed around the bridge model monitor wind field changes in real time, and the feedback module compares the measured wind speed with the target wind speed and outputs the wind speed data. The speed deviation is input to the Transformer neural network; the Transformer neural network dynamically generates a multi-fan array control correction signal based on the error sequence, so that the simulated wind field quickly approaches the target wind field, realizing a high-fidelity simulation of sudden gusts in the canyon; if it is found that the actual wind speed is slightly lower than the target at a certain moment, the control system immediately fine-tunes the fan output to make up for the error; the entire sudden gust process lasts for several seconds, and under the closed-loop control, the wind speed time history in the wind tunnel basically matches the target; through this intelligent simulation system, the sudden gust environment of the canyon where the bridge is located was successfully reproduced in the laboratory, and the wind-induced response data of the bridge model thus has a reliability close to the real conditions.

[0032] Implementation Case 2: Gust Simulation at a Plateau Wind Farm A plateau wind farm frequently encounters severe gusts, and the performance of its wind turbines in unsteady wind fields needs to be evaluated. This invention's system is used to simulate the gust cluster process of this wind farm. First, wind speed monitoring data of the wind farm area is collected. Based on wavelet packet energy entropy, wind field characteristics of gust clusters (such as several consecutive sudden gusts and intermittent gusts) are extracted. The extracted wind field characteristics are then used to train an MLP mapping model to generate multi-fan array control signals, controlling the multi-fan array to generate a simulated gust cluster wind field. The Transformer neural network feedback module dynamically optimizes the multi-fan array control signals according to the changes in wind field error during continuous gusts, achieving dynamic high-fidelity reproduction of multiple gust pulses, providing wind field conditions for verifying the gust resistance performance of the wind turbine. In the experiment, the system controls the fan array to generate multiple strong wind pulses sequentially, combined with intermediate wind speed descent sections, forming a gust cluster flow field. Through real-time feedback correction of the Transformer neural network, each wind speed peak and valley value is accurately reproduced. The experimental results show that the load response of the wind turbine model in gust clusters can be captured experimentally, providing valuable basis for wind turbine design.

[0033] Implementation Case 3: Simulation of strong gusts on power transmission lines in mountainous areas A high-altitude power transmission line may experience conductor icing and vibration instability under extreme wind and low temperature conditions. This invention simulates the wind field effect under this scenario. A historical extreme wind event accompanied by low temperatures is selected as the target wind field, characterized by continuous high-speed winds mixed with short-duration strong gusts. Before the experiment, a multi-fan array mapping model is trained using similar meteorological data to learn the wind speed and tilt angle characteristics of each fan under this abrupt wind field. Then, the target wind field parameters are input into the MLP mapping model to generate multi-fan array control signals. The multi-fan array control system controls the fan array to generate a continuous high-wind background flow field based on the multi-fan array control signals, while simultaneously superimposing short-duration strong gust pulses that appear at random times to simulate the basic wind under gust disturbance. Through the internal and external double-loop feedback optimization of the Transformer neural network, the simulated wind speed pulses are highly consistent with the actual records. The entire wind field effect lasts for tens of seconds, and the model conductor in the wind tunnel experiences a strong gust environment similar to that in the real world. This simulation helps engineers observe the vibration behavior of iced conductors under extreme wind fields in the laboratory, providing a basis for formulating protective measures.

[0034] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0035] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A wind tunnel simulation method for actively controlling abrupt wind field changes in mountainous areas based on deep learning, characterized in that, Includes the following steps: S1: Deploy an array of anemometers in the mountainous target area to obtain the target wind field parameters of the mountainous target area, perform abrupt change identification on the target wind field parameters, and construct abrupt change feature vector; construct a wind field feature vector based on the target wind field parameters; perform principal component analysis (PCA) and mutual information filtering on the wind field feature vector and real-time wind tunnel sensor data to construct an optimal feature subset, and combine the abrupt change feature vector with the optimal feature subset to construct an MLP input vector; S2: Establish a deep learning control model, including a forward mapping model and a feedback optimization model. The forward mapping model adopts a multilayer perceptron (MLP), and the feedback optimization model adopts a time-series feedback model based on a Transformer neural network. The MLP is used to establish the mapping relationship between the target wind field parameters and the control signals of the multi-fan array in the wind tunnel. The Transformer neural network is used to dynamically correct the deviation between the simulated wind field parameters and the target wind field parameters in the wind tunnel test section. S3: Input the MLP input vector into the multilayer perceptron MLP to generate a multi-fan array control signal; S4: The multi-fan array control system operates the multi-fan array according to the multi-fan array control signal to generate a simulated wind field in the wind tunnel test section; S5: Collect real-time wind field data of the simulated wind field through an array of wind speed and direction sensors, and calculate the wind field difference between the real-time wind field data and the target wind field parameters; S6: The Transformer neural network establishes an error evolution law model based on the wind field difference and historical control data, predicts the error evolution trend, and generates a multi-fan array control correction signal based on the error evolution trend; S7: The multi-fan array control system adjusts the speed and tilt angle of the multi-fan array according to the multi-fan array control correction signal; S8: Repeat steps S5 to S7 until the wind field difference meets the preset stability criterion.

2. The wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning according to claim 1, characterized in that, In step S1, the deployment method of the anemometer array is as follows: the mountainous target area is divided into an unencrypted zone and an encrypted zone according to CFD simulation, and the array is deployed in the unencrypted zone according to a preset density. Deploy anemometers of wind speed and direction in the densely populated area according to a preset deployment density. Deploy anemometers and wind direction indicators. < The encryption zone includes: an airflow acceleration zone, a separation zone, and a reattachment zone.

3. The wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning according to claim 1, characterized in that, In step S1, the mutation identification operation includes the following steps: S11: Kalman filtering is used to denoise the target wind field parameters to obtain wind field preprocessing data; S12: Employing a mutation identification algorithm based on wavelet packet energy entropy or variational mode decomposition (VMD), the wind speed sequence in the wind field preprocessing data is decomposed into multiple frequency bands according to a preset frequency band range and number of frequency bands. This detects gust mutation intervals and energy distribution changes, extracts mutation features to form the mutation feature vector, which includes: wind speed amplitude mutation features, wind speed temporal mutation features, wind speed energy distribution features, turbulence features, and spatial correlation features.

4. The wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning according to claim 3, characterized in that, The mutation identification operation introduces an online self-learning mechanism, which automatically adjusts the denoising parameters of the Kalman filter, the preset frequency band range and number of frequency bands, and the parameter composition of the mutation feature vector based on the distribution changes of the target wind field parameters. The wind speed amplitude mutation features include: wind speed value at the start of the mutation, peak wind speed, wind speed amplitude change rate, and wind speed value at the end of the mutation. The wind speed temporal mutation features include: duration of gust mutation, delay time of peak wind speed occurrence relative to the start of the mutation, duration of wind speed rise phase, and duration of wind speed fall phase. The wind speed energy distribution features include: high-frequency energy proportion and extreme values ​​of wavelet packet energy entropy within the mutation interval. The turbulence features include: turbulence intensity within the mutation interval, peak value of the fluctuating wind speed power spectrum in the mutation frequency band, and maximum change in wind angle of attack. The spatial correlation features include: wind speed difference and wind direction deviation between adjacent measuring points within the mutation interval. The parameter composition of the wind field feature vector includes: wind speed component along the horizontal mainstream direction of the target mountain area. Horizontal lateral wind speed component perpendicular to the horizontal mainstream direction Vertical wind speed component perpendicular to the horizontal mainstream direction Parameters including wind angle of attack, wind shear, turbulence intensity, and power spectrum of fluctuating wind speed.

5. The wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning according to claim 1, characterized in that, After each round of wind tunnel experiments, the deep learning control model uses a joint loss function to co-train the multilayer perceptron (MLP) and the Transformer neural network. The control signal output by the MLP serves as the initial input state for the Transformer neural network, which corrects its weight parameters through time-series feedback. The MLP employs a loss function that includes a wind tunnel wind speed distribution error term during the training phase. The joint loss function is expressed as follows: ； ； ； In the formula, For the joint loss function, and These are the loss terms of the Multilayer Perceptron (MLP) and the Transformer Neural Network, respectively. and These are the weight coefficients for the MLP loss term and the Transformer loss term, respectively. These are the control commands output by the multilayer perceptron (MLP). For target control commands, The wind speed distribution is formed after the fan array is driven by the control commands output by the multilayer sensor (MLP). For the target wind speed distribution, The wind field error predicted by the Transformer neural network. This represents the actual wind field error. These are the control commands modified by the Transformer neural network.

6. A wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning, as described in claim 1 or 5, characterized in that... The Transformer neural network incorporates wind tunnel flow field conservation and energy constraints into its attention weight calculation formula, which is expressed as follows: ； ； ； ； In the formula, Here is the formula for calculating the attention weights in a Transformer neural network. , and These are the query vector, key vector, and value vector, respectively. For normalization function, For transpose operation, Let be the dimension of the key vector. and These are the weighting coefficients for the flow field conservation constraint term and the energy constraint term, respectively. and These are respectively the flow field conservation constraint and the energy constraint. This represents the wind speed vector field in the wind tunnel. Let be the divergence of the wind speed vector field in the wind tunnel. The total kinetic energy of the wind field in the wind tunnel. The total kinetic energy of the target wind field air density, This is the space area for the wind tunnel test section. This represents the wind speed component along the horizontal extension direction of the wind tunnel. The horizontal lateral wind speed component is perpendicular to the horizontal extension direction of the wind tunnel. The vertical wind speed component is perpendicular to the horizontal extension direction of the wind tunnel.

7. The wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning according to claim 1, characterized in that, The deep learning control model introduces a control policy knowledge distillation mechanism during the training of the Transformer neural network. This mechanism extracts hidden features from the historical control data and transfers these features to the encoding layer of the current Transformer neural network. The historical control data includes control command sequences and wind field response information recorded during all wind tunnel experiments prior to the current wind tunnel experiment. These hidden features serve as auxiliary information during the training process and participate in the parameter updates of the current Transformer neural network. The deep learning control model pre-configures control policy templates for various typical mountain wind field types, including canyon gusts, mountaintop wind shear, and plateau pulse turbulence. These control policy templates are used to control the multi-fan array to generate the initial wind field in the wind tunnel test section.

8. The wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning according to claim 1, characterized in that, The multi-fan array includes multiple independently controllable fans arranged upstream of the wind tunnel test section. These fans are arranged in a preset two-dimensional grid array. The fans are divided into a core area fan group, a core area peripheral fan group, and an edge area fan group according to a preset ratio. The core area fan group is used to simulate sudden changes in the wind field, the core area peripheral fan group is used for background flow field compensation, and the edge area fan group is used to provide basic wind speed support. Each fan is equipped with an independent servo drive device, which independently adjusts the speed and tilt angle of each fan according to the initial multi-fan array control signal or the corrected multi-fan array control signal.

9. The wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning according to claim 1, characterized in that, The multi-fan array control system includes an adaptive threshold judgment mechanism, which is used to switch the operating mode of the multi-fan array control system according to the changing trends of wind speed amplitude error, wind direction deviation and turbulence intensity error. The operating modes include a sudden change response mode and a steady state maintenance mode.

10. A wind tunnel simulation method for active control of abrupt wind field changes in mountainous areas based on deep learning, as described in claim 1, is characterized in that... The wind field difference includes: wind speed amplitude error, wind direction deviation, and turbulence intensity error; the preset stability criteria include: the wind speed amplitude error is less than a preset wind speed amplitude error threshold, the wind direction deviation is less than a preset wind direction deviation threshold, and the turbulence intensity error is less than a preset turbulence intensity error threshold.

Images