A method and system for gust flow field prediction

Abstract

The application discloses a gust flow field prediction method and system, relates to the technical field of unsteady fluid mechanics, and comprises the following steps: acquiring gust flow field data and a gust generator control parameter under different working conditions, and constructing a training data set and a test data set; constructing a gust flow field prediction model, training the gust flow field prediction model based on the training data set, and obtaining a trained gust flow field prediction model; inputting the test data set into the trained gust flow field prediction model to obtain gust velocity field prediction data; and comparing and evaluating the gust velocity field prediction data with original gust velocity field data in the test data set to obtain an evaluation result. The application can realize extrapolation prediction of the evolution characteristics of the unsteady gust flow field, and improve the accuracy and stability of gust flow field prediction.

Description

Technical Field

[0001] This application relates to the field of unsteady fluid mechanics, and in particular to a method and system for predicting gust flow fields. Background Technology

[0002] In recent years, with the rapid development of computer science and artificial intelligence technologies, data-driven methods have been gradually applied to wind speed prediction and unsteady fluid mechanics. In fluid mechanics, addressing the issues of large computational loads and high numerical simulation costs associated with high-dimensional flow field data, related technologies have proposed using deep learning models such as autoencoders to perform low-dimensional characterization of high-dimensional flow fields. This allows for the depiction of complex flow structure features through latent variables, providing a new technical path for data compression and rapid reconstruction.

[0003] However, most existing studies focus on flow field modeling methods driven by a single data source. In the feature extraction stage, they typically rely solely on the velocity field image itself for encoding, lacking effective utilization of local waveform information at measurement points during gust propagation. This results in the obtained potential low-dimensional features failing to fully reflect the dynamic evolution of gust propagation. Furthermore, during model training, existing methods often employ fixed-weight multi-objective loss functions to jointly optimize flow field reconstruction errors and auxiliary physical information errors. Since the importance of different loss terms to model learning varies between the early and later stages of training, the fixed-weight approach can easily lead to instability in the training process, even causing problems such as model convergence difficulties or premature influence from auxiliary constraints, ultimately resulting in low accuracy and stability in gust flow field prediction.

[0004] Therefore, it is urgent to design a method for predicting gust flow fields in order to extrapolate and predict the evolution characteristics of unsteady gust flow fields and improve the accuracy and stability of gust flow field prediction. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for predicting gust flow fields, which can extrapolate and predict the evolution characteristics of unsteady gust flow fields, thereby improving the accuracy and stability of gust flow field prediction.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for predicting gust flow fields, including: Acquire gust flow field data and gust generator control parameters under different operating conditions, and construct training datasets and test datasets; the gust flow field data includes raw gust velocity field data and corresponding measurement point waveform information.

[0007] A gust flow field prediction model is constructed and trained based on the training dataset to obtain a trained gust flow field prediction model. The gust flow field prediction model includes an enhanced autoencoder model that integrates measurement point waveform information. The enhanced autoencoder model that integrates measurement point waveform information adopts an automatic stage switching mechanism and dynamic weights for staged training. It integrates measurement point waveform information including measurement point waveform velocity information and spatial position information on the basis of the autoencoder. By constructing a position-aware attention fusion mechanism, the adaptive fusion of latent features and measurement point waveform features is achieved.

[0008] The test dataset is input into the trained gust flow field prediction model to obtain gust velocity field prediction data.

[0009] The predicted gust velocity field data is compared and evaluated with the original gust velocity field data in the test dataset to obtain the evaluation results.

[0010] Secondly, this application provides a gust flow field prediction system, which is used to implement the gust flow field prediction method described in the first aspect, and the gust flow field prediction system includes: The data acquisition and processing module is used to acquire gust flow field data and gust generator control parameters under different working conditions, and to construct training datasets and test datasets; the gust flow field data includes raw gust velocity field data and corresponding measurement point waveform information.

[0011] The model building and training module is used to build a gust flow field prediction model and train the gust flow field prediction model based on the training dataset to obtain a trained gust flow field prediction model. The gust flow field prediction model includes an enhanced autoencoder model that integrates measurement point waveform information. The enhanced autoencoder model that integrates measurement point waveform information adopts an automatic stage switching mechanism and dynamic weights for staged training. It integrates measurement point waveform information including measurement point waveform velocity information and spatial position information on the basis of the autoencoder. By constructing a position-aware attention fusion mechanism, it achieves adaptive fusion of latent features and measurement point waveform features.

[0012] The testing module is used to input the test dataset into the trained gust flow field prediction model to obtain gust velocity field prediction data.

[0013] The prediction and evaluation module is used to compare and evaluate the predicted gust velocity field data with the original gust velocity field data in the test dataset to obtain the evaluation result.

[0014] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and system for predicting gust flow fields. By collecting gust flow field data and gust generator control parameters under different operating conditions, it solves the technical shortcomings of traditional flow field prediction methods, such as single data source and insufficient extraction of local evolution features, and achieves deep fusion representation of gust spatial structure and dynamic waveform features. By constructing a gust flow field prediction model, it overcomes the problems of discontinuous prediction of periodic gust boundaries and high computational complexity of high-dimensional flow fields, and improves the extrapolation prediction capability of the model under unknown operating conditions. The enhanced autoencoder model that integrates waveform information from measurement points adopts an automatic stage switching machine. The proposed phased training method allows for dynamic adjustment of the loss function weights, improving upon the drawbacks of unstable convergence and difficulty in model convergence associated with fixed-weight training methods, and optimizing the model training process. By applying and testing the trained gust flow field prediction model, it addresses the technical challenges of poor feature representation capabilities, insufficient training stability, and weak generalization ability with small samples found in existing gust flow field prediction models. It enables extrapolation prediction of the evolution characteristics of unsteady gust flow fields, improving the accuracy and stability of gust flow field prediction. Furthermore, the method is highly versatile and adaptable to various flow field prediction scenarios, expanding its engineering application scope. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 An application environment diagram of a gust flow field prediction method provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for predicting gust flow fields according to an embodiment of this application; Figure 3 This is a schematic diagram showing the probability density distribution of the structural similarity index between the predicted gust velocity field data and the original gust velocity field data. Figure 4 This is a schematic diagram showing the predicted gust flow field results at different times; among them... Figure 4 (a) in the middle is θ =12°, k =0.8, t / T The original result when =0.25; Figure 4 (b) in the middle is θ =12°, k =0.8, t / T Prediction results when = 0.25; Figure 4 (c) in the middle is θ =12°,k =0.8, t / T The original result when =0.50; Figure 4 (d) in the middle is θ =12°, k =0.8, t / T Prediction results when =0.50; Figure 5 A flowchart illustrating a gust flow field prediction method provided in another embodiment of this application; Figure 6 A schematic diagram of the functional modules of a gust flow field prediction system provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In atmospheric boundary layer wind speed prediction, deep learning models can extract complex nonlinear temporal features from historical data, achieving some progress in short-term wind speed prediction accuracy. Existing research largely focuses on statistically significant continuous turbulence prediction or flow field reconstruction. For gust evolution problems with significant unsteady and abrupt characteristics, a mature and universal prediction framework is still lacking. Gust flow fields typically exhibit high-dimensional, multivariate spatial distribution data. Directly predicting and modeling high-dimensional velocity fields is not only computationally complex but also prone to the curse of dimensionality, leading to reduced model training efficiency and decreased prediction stability. With a limited sample size, traditional high-dimensional modeling methods often struggle to fully learn the key feature structures implicit in the velocity field.

[0019] Existing autoencoder (AE) models typically employ simple feature splicing or uniform weighting for information fusion in different spatial locations or under different flow conditions. This makes it difficult to adaptively adjust the importance of different information sources according to changes in spatial location, thus limiting the autoencoder model's ability to express complex unsteady gust structures.

[0020] In the training process of autoencoder models, traditional methods typically lack a dynamic training mechanism that can automatically adjust the loss weights based on the training state. This makes it difficult to gradually introduce physical constraint information while ensuring the accuracy of flow field reconstruction, thus limiting the overall convergence efficiency and prediction stability of the autoencoder model. On the other hand, existing data-driven prediction models are generally highly "black box" in nature, lacking auxiliary constraint mechanisms related to the actual gust waveform evolution process, making it difficult to establish a clear connection between latent variables and actual physical processes. When the number of training samples is limited or the range of control parameters changes, the model often exhibits problems such as decreased prediction accuracy, insufficient extrapolation ability, and reduced stability, making it difficult to meet the actual requirements of engineering applications for model reliability and interpretability.

[0021] To address the aforementioned technical challenges, this application proposes a gust flow field prediction method. By constructing a position-aware attention feature fusion mechanism, a periodically embedded sub-target latent low-dimensional feature gust flow field prediction model, waveform-enhanced feature reconstruction, and an automatic stage-switching dynamic training strategy, this method enables the extrapolation prediction of the evolution characteristics of unsteady gust flow fields. This improves the stability, generalization ability, and physical interpretability of the gust flow field prediction model under small sample conditions, thereby providing reliable data support for engineering applications such as aircraft dynamic response analysis, structural load assessment, and active flow control.

[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] The gust flow field prediction method provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send gust flow field data and gust generator control parameters under different operating conditions to server 104. After receiving the gust flow field data and gust generator control parameters under different operating conditions, server 104 constructs a training dataset and a test dataset based on the gust flow field data and gust generator control parameters under different operating conditions; constructs a gust flow field prediction model and trains the gust flow field prediction model based on the training dataset to obtain a trained gust flow field prediction model; inputs the test dataset into the trained gust flow field prediction model to obtain gust velocity field prediction data; and compares and evaluates the gust velocity field prediction data with the original gust velocity field data in the test dataset to obtain an evaluation result. Server 104 can feed back the obtained evaluation results to terminal 102. Furthermore, in some embodiments, the gust flow field prediction method can also be implemented independently by server 104 or terminal 102. For example, terminal 102 can directly perform gust flow field prediction and comparison evaluation processing on gust flow field data and gust generator control parameters under different operating conditions. Alternatively, server 104 can obtain gust flow field data and gust generator control parameters under different operating conditions from the data storage system and perform gust flow field prediction and comparison evaluation processing on the gust flow field data and gust generator control parameters under different operating conditions.

[0024] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0025] In one exemplary embodiment, such as Figure 2 As shown, a method for predicting gust flow fields is provided. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 204.

[0026] Step 201: Obtain gust flow field data and gust generator control parameters under different working conditions, and construct training dataset and test dataset; the gust flow field data includes raw gust velocity field data and corresponding measurement point waveform information.

[0027] It should be noted that the control parameters of the gust generator include the maximum angle of motion and the dimensionless reduction frequency of the gust generator. The waveform information at the measuring point includes the waveform velocity information and spatial position information at the measuring point.

[0028] Step 202: Construct a gust flow field prediction model and train the gust flow field prediction model based on the training dataset to obtain a trained gust flow field prediction model. The gust flow field prediction model includes an enhanced autoencoder model that integrates measurement point waveform information. The enhanced autoencoder model that integrates measurement point waveform information adopts an automatic stage switching mechanism and dynamic weights for staged training. It integrates measurement point waveform information including measurement point waveform velocity information and spatial position information on the basis of the autoencoder. By constructing a position-aware attention fusion mechanism, it achieves adaptive fusion of latent features and measurement point waveform features.

[0029] It should be noted that the gust flow field prediction model also includes a dual-channel independent random forest regression prediction model; the enhanced autoencoder model that fuses measurement point waveform information includes an encoding module and a decoding module; the encoding module is used to determine the fused latent low-dimensional features based on the training dataset; the dual-channel independent random forest regression prediction model is used to determine the predicted latent low-dimensional features based on the training dataset and the fused latent low-dimensional features; the decoding module is used to determine the gust velocity field prediction data based on the predicted latent low-dimensional features. The encoding module of the enhanced autoencoder model that fuses measurement point waveform information includes an autoencoder and a measurement point waveform encoding module (i.e., a measurement point waveform encoding network).

[0030] It should be noted that the gust velocity field prediction data provides a forecast of the spatial state of unsteady flow fields, reflecting the velocity distribution of gusts at different locations within the flow field. The dynamic weights are calculated based on an exponentially increasing weighting function.

[0031] Step 203: Input the test dataset into the trained gust flow field prediction model to obtain gust velocity field prediction data.

[0032] Step 204: Compare and evaluate the predicted gust velocity field data with the original gust velocity field data in the test dataset to obtain the evaluation result.

[0033] It should be noted that the evaluation results are used to determine whether the prediction results of the gust flow field prediction model have good prediction accuracy, especially whether it can accurately predict the peak evolution characteristics of the gust velocity field.

[0034] In this embodiment, 17 sets of gust flow field data under different working conditions are used to train and validate the gust flow field prediction model. Each set of working conditions includes... The original data of 100 gust velocity fields within the range of 0.01 to 1 are provided. The velocity field of each original gust velocity field is a 1×240×240 matrix, representing the dimensionless x-axis values. x / c Dimensionless ordinate value y / c The dimensionless velocity waveform measurement value at the corresponding position v / U ∞ .in, Time with dimensions; For gust cycles; x for x Direction is a spatial position with dimensions; y for y Direction is a spatially quantifiable position; c For the chord length of the gust generator; v This refers to the vertical gust speed. U ∞ This represents the incoming flow velocity. In addition, the gust flow field data also includes single-point velocity waveform measurements at a specific moment. v / U ∞ The measurement location was between the two gust generator blades and 3.4 meters from the trailing edge of the blades. c Place.

[0035] Table 1. Division of Test Set and Training Set Operating Conditions

[0036] Among them, the gust generator control parameters corresponding to each operating condition θ and k As shown in Table 1, θ This represents the maximum angle of motion of the gust generator; k The frequency was reduced to dimensionless values. To evaluate the performance of the gust flow field prediction model, 13 sets of sin gust flow field data with good homogeneity were used as the training dataset, and the remaining 4 sets of gust flow field data with strong nonlinearity and poor homogeneity were used as the test dataset. In addition, 20% of the data from the 4 operating conditions in the test dataset were randomly selected as known data and added to the test dataset. This study effectively verifies the ability of the embodiments of this application to achieve high-fidelity and high-efficiency extrapolation prediction of the evolution characteristics of unsteady gust flow fields using limited sparse gust experimental data.

[0037] By implementing steps 201 to 204 above, the embodiments of this application can achieve extrapolation prediction of the evolution characteristics of unsteady gust flow fields, improving the accuracy and stability of gust flow field prediction. Furthermore, the method of the embodiments of this application can be extended to the evolution prediction of various flow parameters such as temperature field, pressure field, and vorticity field, and can be applied to different types of unsteady flow problems, thereby significantly expanding the applicability and engineering application value of the method.

[0038] In another exemplary embodiment of this application, for high-dimensional gust velocity field raw data, directly using random forest prediction will result in problems such as loss of spatial correlation, output dimension explosion, and mismatch between input features and output dimensions. To address these issues, this embodiment constructs an enhanced autoencoder model that fuses measurement point waveform information, based on the traditional autoencoder model using convolutional neural networks (CNN) and multilayer perceptrons (MLP). This model is used to achieve efficient compression and low-dimensional feature extraction of gust flow field data. The step of determining the fused potential low-dimensional features based on the training dataset in the encoding module of the enhanced autoencoder model is replaced by steps 301-302: Step 301: Input the original data of the gust velocity field and the waveform information of the measuring point into the encoding module of the enhanced autoencoder model that fuses the waveform information of the measuring point to obtain the latent features and the waveform features of the measuring point.

[0039] Specifically, in this application embodiment, the input of the encoding module of the enhanced autoencoder model that integrates measurement point waveform information includes the raw data of the gust velocity field and its corresponding measurement point waveform information, wherein the raw data of the gust velocity field at a certain moment is single-channel two-dimensional data: ;in, This is the raw data of the gust velocity field at a certain moment; The space is real numbers, and the superscript indicates the original data of each gust velocity field. The matrix.

[0040] Simultaneously, waveform information from the measurement points is introduced, including dimensionless time variables. Measurement point velocity value and the spatial coordinates of the measuring point ( ).in, The x-coordinate of the spatial location of the measuring point; The spatial coordinates of the measurement points are used to enhance the network's ability to learn spatial structures. The spatial location information of the measurement points is normalized and mapped to a standard interval, thereby ensuring the stability of the training of the enhanced autoencoder model that integrates the waveform information of the measurement points at different spatial scales.

[0041] The raw data of the gust velocity field is first input into the encoder module of the enhanced autoencoder model that integrates the waveform information of the measurement points for data compression. The encoder compresses the spatial dimension and extracts global features step by step through multiple convolution and pooling operations in the convolutional neural network to obtain the intermediate layer. X conv The network is constructed as a 4×12×12 network, which can be modified according to the size of the dataset. Then, a fully connected network of multilayer perceptrons is used to further reduce the dimensionality and obtain the latent features.

[0042] The expression for the latent feature (i.e., the encoding process of the raw data of the gust velocity field) is as follows: (1); in, For the encoding function, the output dimension can be set in the hidden layer. To represent latent features, taking two dimensions as an example, latent features can be represented as follows: =[ 1, 2] T (in, 1 represents the abscissa value of the latent feature in space; 2 represents the ordinate value of the latent feature in space; [·] T (This is a vector transpose), representing the two-dimensional latent features obtained after dimensionality reduction of the original gust velocity field data through the encoding module of an enhanced autoencoder model that integrates waveform information from measurement points. The compressed feature dimension can be adjusted according to the features of the original training dataset.

[0043] To enhance the ability of the enhanced autoencoder model that integrates measurement point waveform information to characterize the dynamic evolution of gusts, this application further constructs a measurement point waveform encoding network to extract features from the measurement point waveform information and obtain measurement point waveform features.

[0044] The expression for the waveform characteristics of the measurement point (i.e., the encoding process of the waveform information of the measurement point) is as follows: (2); in, Indicates the waveform characteristics of the measuring point; The waveform encoding function at the measurement point extracts phase changes and local velocity evolution features from the waveform information at the measurement point through multi-layer nonlinear mapping, and outputs the waveform features at the measurement point. =[ 1, 2] T (in, 1 represents the abscissa value of the waveform characteristics of the measuring point in space; 2 represents the vertical coordinate of the waveform feature at the measurement point in space. This feature can effectively supplement the local gust propagation information that is difficult for AE to obtain directly.

[0045] Step 302: A position-aware attention fusion mechanism is used to perform weighted fusion of the latent features and the waveform features of the measurement points to obtain the fused latent low-dimensional features.

[0046] Specifically, after obtaining the latent features and the waveform features of the measurement points, this application proposes a position-aware attention fusion mechanism to achieve adaptive weighted fusion of multi-source features and obtain the fused latent low-dimensional features.

[0047] The expression for the fused latent low-dimensional features is: (3); in, This represents the potential low-dimensional features after fusion; For attention weights.

[0048] The formula for calculating the attention weight is: (4); in, , and The weight matrix is ​​trainable. Spatial location of the measuring point; symbol This indicates element-wise multiplication.

[0049] By introducing offset information (i.e., spatial location information) of the measuring points' spatial positions, this fusion method can dynamically adjust the contribution weight of the measuring point waveform information according to different gust flow field states, making the fusion result more consistent with the spatial physical laws of gust propagation and avoiding the information redundancy problem caused by the traditional simple feature splicing method. Finally, the fused potential low-dimensional feature is defined as... =[ 1, 2] T (in, 1 represents the abscissa value of the fused potential low-dimensional feature in space; 2 represents the ordinate value of the fused potential low-dimensional feature in space. This fused potential low-dimensional feature simultaneously contains information on the spatial structure of the flow field, dynamic information on the waveform velocity of the measurement point, and information on spatial position constraints, providing high-quality input features for the subsequent establishment of a dual-channel independent random forest regression prediction model.

[0050] Unlike traditional autoencoders that only use images for encoding, the embodiments of this application introduce measurement point waveform information, including measurement point waveform velocity information and spatial location information, corresponding to the original data of the gust velocity field during the encoding process. The adaptive fusion of multi-source features is achieved through a position-aware attention fusion mechanism, thereby significantly improving the characterization ability of potential low-dimensional features of gust evolution characteristics.

[0051] In an exemplary embodiment of this application, in order to predict the complete evolution law of potential low-dimensional features, this embodiment proposes a method for predicting potential low-dimensional features by combining periodic feature embedding and sub-target dual-channel independent random forest (RF) regression prediction model. Based on known gust generator control parameters and the fused potential low-dimensional features corresponding to the training dataset, the step of determining the predicted potential low-dimensional features using the dual-channel independent random forest regression prediction model based on the training dataset and the fused potential low-dimensional features is replaced by steps 401-403: Step 401: Calculate the dimensionless time parameter in the original data of the gust velocity field to obtain the periodic embedding feature.

[0052] Step 402: Based on the periodic embedding features and the gust generator control parameters, an input feature vector is constructed.

[0053] Specifically, to address the issue that continuous sin gusts exhibit significant periodicity and phase dependence, a periodic function mapping mechanism is introduced, thereby significantly enhancing the ability of the dual-channel independent random forest regression prediction model to express phase changes and its generalization performance.

[0054] The input feature vector is denoted as A =[ θ , k sin(2π) t / T ),cos(2π t / T )] T The maximum angle of motion of the gust generator θ Dimensionless reduced frequency k and for dimensionless time variables t / T The periodic embedding feature obtained by further calculation of the parameters is sin(2π) t / T ) and cos(2π t / TThe output target is the fused latent low-dimensional features extracted by the encoder in the encoding module of the enhanced autoencoder model that integrates the waveform information of the measurement points. Among them, the periodic embedding feature transforms the original time axis from linear space to phase space on the unit circle, enabling the dual-channel independent random forest regression prediction model to eliminate the problem of discontinuity at the time boundary, while enhancing its ability to express the response to periodic gusts.

[0055] Step 403: Input the input feature vector into the dual-channel independent random forest regression prediction model, and use the fused potential low-dimensional features as the output target to obtain the predicted potential low-dimensional features.

[0056] Specifically, as an efficient ensemble learning algorithm, random forests, by constructing multiple decision trees and integrating the prediction results of each decision tree, can not only effectively capture the nonlinear relationships between variables, but also significantly reduce the overfitting risk of dual-channel independent random forest regression prediction models.

[0057] This application's embodiments employ a dual-channel independent random forest regression prediction model, specifically targeting the fused latent low-dimensional features separately. 1 and 2 is considered as an independent prediction target, resulting in a dual-channel independent random forest regression prediction model.

[0058] The expression for the predicted latent low-dimensional features is: (5); (6); in, This represents the latent low-dimensional features of the prediction obtained from the first prediction channel; This represents the potential low-dimensional features of the predictions obtained from the second prediction channel; This represents the expression for the two-channel independent random forest regression prediction model for the first prediction channel; This represents the expression for the two-channel independent random forest regression prediction model for the second prediction channel.

[0059] Each dual-channel independent random forest regression prediction model is composed of The system consists of several regression trees. During training, each tree uses a bootstrap sampling method to extract subsamples from the training dataset, and randomly selects a subset of features for each node split. Taking the predicted latent low-dimensional features obtained from the first prediction channel as an example, given the input feature vector... A The predictions of a random forest are obtained by averaging the outputs of all trees.

[0060] The expression for the dual-channel independent random forest regression prediction model of the first prediction channel is: (7); in, For the first The prediction results of the decision tree; The number of decision trees; For decision tree indexing.

[0061] The goal of each decision tree is to minimize the mean squared error in the leaf nodes. The final ensemble result can effectively smooth the high variance prediction of a single tree and improve the generalization ability of the dual-channel independent random forest regression prediction model.

[0062] In order to obtain the optimal hyperparameter combination for the dual-channel independent random forest regression prediction model, this application adopts a combination of grid search and cross-validation to determine the expression of the hyperparameter grid space and the expression of the optimal hyperparameter selection target.

[0063] The expression for the hyperparameter grid space is: (8); in, For hyperparameter grid space; The maximum depth of the decision tree; The minimum number of samples required for node splitting; This represents the minimum number of samples for a leaf node.

[0064] The expression for the objective of selecting the optimal hyperparameters is: (9); in, Selecting an objective for optimal hyperparameters; For hyperparameter combinations; For hyperparameter grid space A certain hyperparameter combination The calculated loss function; This represents the cross-validation error.

[0065] The dual-channel independent random forest regression prediction model uses mean squared error as the loss function during training. The expression for the loss function is: (10); in, Indicates the first i The first actual latent low-dimensional feature of each sample Indicates the first i The first predicted latent low-dimensional feature of each sample; N The number of samples; i For sample index.

[0066] Finally, use S The corresponding optimal dual-channel independent random forest regression prediction model is retrained on the complete training dataset.

[0067] In another exemplary embodiment of this application, the decoder in the decoding module of the enhanced autoencoder model that fuses measurement point waveform information will predict the latent low-dimensional features. The data is mapped back to the intermediate layer through a fully connected layer, and then restored to the original gust velocity field size of 240×240 through layer-by-layer upsampling and convolution.

[0068] The expression for the predicted gust velocity field data is: (11); in, This is the predicted data for gust speed fields; The decoder function consists of a fully connected network and a multi-layer deconvolution structure (upsampling + convolution). Its goal is to restore the details of the input gust velocity field as much as possible. This process satisfies the following constraints.

[0069] The expression for the constraint condition is: (12); in, For autoencoder functions, including encoder functions With decoder function ; The weights, determined by the Adam optimizer, are optimized to minimize the cost (loss) function of the reconstruction error. Ultimately, the enhanced autoencoder model, which fuses the waveform information from the measurement points, achieves compression and feature extraction while preserving the main information of the gust velocity field. By improving the structure of the multilayer perceptron (MLP) connection layers, the original 4-layer network structure is increased to 6 layers. This allows the fully connected layers to gradually reduce dimensionality and upsample under a denser grid, retaining more effective information and achieving better compression and reconstruction results.

[0070] In addition, consider the encoded vector (i.e., the latent low-dimensional features after fusion). Based on the physical information that may be contained in the nonlinear gust waveform evolution, a gust waveform enhancement network (i.e., gust response prediction branch) was constructed, which is a velocity waveform decoder composed of a three-layer feedforward neural network structure.

[0071] The formula for calculating the predicted velocity waveform value is as follows: (13); in, This is a speed waveform decoding function; For the predicted value of the velocity waveform, y = v / U ∞ The measured value is the velocity waveform. yes The prediction target.

[0072] The velocity waveform prediction value is used to provide a physically meaningful auxiliary loss when constructing the loss function of the enhanced autoencoder model that integrates the waveform information of the measurement points, making the learning of the enhanced autoencoder model that integrates the waveform information of the measurement points more physically interpretable.

[0073] In one exemplary embodiment of this application, to improve the physical interpretability and reconstruction accuracy of potential low-dimensional features during the training phase of the gust flow field prediction model, this embodiment employs a multi-objective joint optimization strategy to collaboratively train the flow field reconstruction error and the gust waveform error. Unlike traditional loss functions using fixed weights, this embodiment proposes a dynamic training method based on automatic stage switching and exponentially increasing weights (i.e., dynamic weights). By adaptively adjusting the loss function weights, it achieves a progressive optimization process from data-driven learning to physical constraint learning for the gust flow field prediction model, thereby improving the stability and convergence efficiency of the model.

[0074] The automatic stage switching mechanism automatically switches between the first training stage and the second training stage based on the preset ratio between the flow field reconstruction loss and the initial flow field reconstruction loss.

[0075] In the early stage of training (i.e., the first training phase), the enhanced autoencoder model that integrates the waveform information of the measurement points takes the flow field reconstruction error as the optimization target. The encoding module prioritizes learning the dominant spatial structure of the gust flow field to obtain the flow field reconstruction loss.

[0076] The formula for calculating the flow field reconstruction loss is: (14); in, This represents the loss due to flow field reconstruction.

[0077] In the next training phase (i.e., the second training phase), the enhanced autoencoder model that integrates the waveform information from the measurement points introduces a gust response prediction branch to determine the auxiliary loss. An exponentially increasing weight function is used to weight the flow field reconstruction loss and the auxiliary loss to determine the improved total loss. The gust response prediction branch is a branch of the decoding module in the enhanced autoencoder model that integrates the waveform information from the measurement points. The gust response prediction branch is used to calculate the predicted velocity waveform value based on the fused potential low-dimensional features. The auxiliary loss represents the error loss between the predicted velocity waveform value and the actual measured velocity waveform value.

[0078] The introduction of a gust response prediction branch serves as an auxiliary loss to guide the encoder in learning a more physically interpretable representation of potential variables.

[0079] The formula for calculating the auxiliary loss is: (15); in, To mitigate losses.

[0080] In an exemplary embodiment of this application, the automatic stage switching mechanism automatically switches between the first training stage and the second training stage based on a preset ratio between the flow field reconstruction loss and the initial flow field reconstruction loss, including the following steps 501 to 502.

[0081] Step 501: When the flow field reconstruction loss is greater than a preset ratio of the initial flow field reconstruction loss, the enhanced autoencoder model that fuses the waveform information of the measurement points enters the first training stage, and the training loss of the enhanced autoencoder model that fuses the waveform information of the measurement points is the flow field reconstruction loss.

[0082] Step 502: When the flow field reconstruction loss is less than or equal to a preset proportion of the initial flow field reconstruction loss, the enhanced autoencoder model that fuses the measurement point waveform information automatically switches from the first training stage to the second training stage, and the training loss of the enhanced autoencoder model that fuses the measurement point waveform information is the improved total loss.

[0083] Specifically, during the training process, this embodiment of the application employs an automatic stage switching mechanism. When the flow field reconstruction loss decreases to a certain percentage threshold of the initial flow field reconstruction loss, the second training stage is automatically triggered, allowing the enhanced autoencoder model that fuses the waveform information of the measurement points to gradually introduce gust response constraints.

[0084] The expression for the switching condition of the automatic phase switching mechanism is: (16); in, This represents the initial flow field reconstruction loss. This is the threshold coefficient for phase switching. Adjustments can be made according to training needs; select appropriate methods during validation. =0.5. During the first training phase, only the flow field reconstruction loss is used; that is, the training loss of the enhanced autoencoder model that integrates measurement point waveform information is the flow field reconstruction loss. This phase is mainly used to establish a stable flow field structure encoding capability, avoiding premature introduction of multi-objective constraints that could lead to training instability. After the phase switching conditions are met, the second training phase begins, introducing an auxiliary loss and using an exponentially increasing weight function to gradually enhance its influence, allowing the model to smoothly transition to a multi-objective joint training state.

[0085] The expression for the improved total loss is as follows: (17); in, The total loss after the improvement; It is an exponentially increasing weighting function.

[0086] The expression for the exponentially increasing weighting function is: (18); in, The maximum weight can be adjusted according to actual training needs, and can be selected in method validation. =0.08; This refers to the number of training rounds after entering the second training phase. It is the time constant for exponential growth.

[0087] To further improve training adaptability, this application proposes a phase-related time constant adjustment mechanism to dynamically adjust parameters based on the start time of the second phase. .

[0088] The expression for the exponential growth time constant is: (19); in, This represents the total number of training rounds. This is the upper limit of the time constant, used in method verification. =15; This is the lower limit of the time constant, used in method verification. =5.

[0089] Through the stage-related time constant adjustment mechanism, when the enhanced autoencoder model that integrates measurement point waveform information enters the second training stage earlier, the growth of the exponentially increasing weights calculated by the exponentially increasing weight function tends to be gradual, thus ensuring training stability; when the enhanced autoencoder model that integrates measurement point waveform information enters the second training stage later, the growth rate of the exponentially increasing weights calculated by the exponentially increasing weight function is accelerated, enabling the auxiliary loss to quickly play a constraining role, thereby improving the overall convergence efficiency.

[0090] Finally, to enhance the ability of the gust flow field prediction model to represent positive and negative gust changes, this embodiment employs a nonlinear activation function for feature mapping in the network structure. This allows the latent space to simultaneously represent gust disturbances from different directions, improving the adaptability of the gust flow field prediction model to complex unsteady flows. Through the aforementioned dynamic stage training strategy, this embodiment can gradually introduce physical constraint information while ensuring the accuracy of flow field reconstruction. This effectively avoids the training instability and weight selection difficulties inherent in traditional fixed-weight methods, significantly improving the convergence speed, stability, and prediction accuracy of the gust flow field prediction model.

[0091] This application provides an exemplary embodiment for verifying the effectiveness of the method described in this embodiment. The division of the test dataset and training dataset is consistent with Table 1, and 20% of the distorted velocity field data is randomly measured as known sparse data. The gust flow field evolution prediction results obtained according to the method of this embodiment are as follows: Figure 3 and Figure 4 As shown. The structural similarity index (SSI) is used. SSIM This is used to evaluate the similarity between the predicted gust flow field and the original gust flow field.

[0092] Figure 3 The average structural similarity index between the predicted gust velocity field data and the original gust velocity field data is 0.749, and the probability density distribution of the structural similarity index is mostly concentrated between 0.7 and 0.9, indicating good overall prediction performance. The predicted gust flow field results at different times are shown below. Figure 4 As shown, where, Figure 4 (a) in the middle is θ =12°, k =0.8, t / T The original result when =0.25; Figure 4 (b) in the middle is θ =12°, k =0.8, t / T Prediction results when = 0.25; Figure 4 (c) in the middle is θ =12°, k =0.8, t / T The original result when =0.50; Figure 4 (d) in the middle is θ =12°, k =0.8, t / T The prediction result when =0.50. t / T When the coefficient of performance (COP) is 0.25, the gust velocity field is near the peak velocity, and the structural similarity index of the predicted result is 0.851, indicating a high degree of similarity between the predicted result and the original result. t / T=At 0.50, this moment corresponds to the equilibrium position of the gust waveform. The gust velocity field varies near 0, resulting in a larger relative error. The structural similarity index of the prediction result is 0.591. The results show that the gust flow field prediction model proposed in this application has good overall prediction accuracy, especially in accurately predicting the peak evolution characteristics of the gust velocity field.

[0093] Another embodiment of this application provides a method for predicting gust flow fields, such as... Figure 5 As shown, the first step is to prepare the dataset (i.e., build the training dataset and the test dataset) and obtain the parameters affecting gusts. θ, k, (i.e., gust generator control parameters), gust velocity field in 1×240×240 matrix form (i.e., raw gust velocity field data). And including location (i.e., spatial location of the measuring point). Time (i.e., dimensionless time variable) Speed ​​(i.e., the speed value at the measuring point) The measurement point waveform data (i.e., measurement point waveform information) is then fused to extract gust flow field features. The raw gust velocity field data is encoded using a convolutional neural network of an autoencoder and a multilayer perceptron to obtain low-dimensional features (i.e., latent features). The measuring point waveform encoding module encodes the measuring point time and velocity information (i.e., measuring point waveform information) to obtain waveform features (i.e., measuring point waveform features) with the same dimension as the gust velocity field characteristics (i.e., potential characteristics). Then, attention weights are calculated through a location-aware attention fusion mechanism. The two types of features are adaptively weighted and fused to obtain the fused features (i.e., the latent low-dimensional features after fusion). Subsequently, based on periodic embedding and a sub-objective random forest model (i.e., a dual-channel independent random forest regression prediction model), low-dimensional features (i.e., latent low-dimensional features) are predicted to construct the input feature vector. A =[ θ , k sin(2π) t / T ),cos(2π t / T )] T It employs random forest combined with sub-objective prediction, hyperparameter grid search, and cross-validation, while also using... =[ 1, 2] T To achieve the output target, the predicted latent low-dimensional features are obtained; then, the predicted features are reconstructed using a waveform enhancement decoder, utilizing the decoder of the autoencoder (i.e., the decoding module of the enhanced autoencoder model that fuses measurement point waveform information). The predicted low-dimensional features (i.e., the predicted latent low-dimensional features) are reconstructed into high-dimensional flow field prediction data, and a waveform enhancement network (i.e., a gust waveform enhancement network) is used to output the velocity waveform prediction value based on the fused latent low-dimensional features. Finally, during the phased dynamic model training, Phase 1 (i.e., the first training phase) uses flow field reconstruction loss. To optimize the training of the autoencoder, Phase 2 (the second training phase) introduces exponentially increasing weights and a dynamically adjusted time constant (i.e., an exponentially growing time constant). The improved total loss is then calculated by combining the flow reconstruction loss and waveform loss (i.e., auxiliary loss). Among them, the automatic switching between the first training phase and the second training phase needs to meet the threshold of the automatic (phase) switching mechanism.

[0094] The enhanced autoencoder model proposed in this application, which integrates measurement point waveform information, introduces measurement point waveform information and spatial location information on the basis of traditional autoencoders. By constructing a position-aware attention fusion mechanism, it achieves adaptive fusion of latent features and measurement point waveform features. This method first uses a convolutional neural network to extract spatial features from high-dimensional gust velocity field data to obtain latent features. Then, a measurement point waveform encoding module is constructed to perform nonlinear mapping on the measurement point time and velocity information in the measurement point waveform information to extract the measurement point waveform features of gust propagation. Furthermore, the two types of features are weighted and fused through the position-aware attention mechanism to obtain a fused latent low-dimensional feature that simultaneously contains spatial structure information, dynamic evolution information, and spatial location constraint information. Through this multi-source information fusion method, this application effectively solves the problem of insufficient local evolution feature representation caused by traditional autoencoders relying solely on image information, improves the ability of latent features to represent the evolution law of unsteady gusts, and thus significantly improves the accuracy of flow field reconstruction and the reliability of prediction.

[0095] This application further proposes a method for predicting latent low-dimensional features in a dual-channel independent random forest regression prediction model based on periodic embedding and sub-objectives. Addressing the issue of significant periodicity and phase dependence in flows such as continuous sin gusts, a periodic function embedding mechanism is introduced to map the time variable to a periodic phase space, thereby enhancing the model's ability to express periodic variations. Simultaneously, this application employs a dual-channel independent random forest regression prediction model, using the fused latent low-dimensional features extracted by an enhanced autoencoder model that integrates waveform information from measurement points as independent prediction targets for separate modeling. Furthermore, grid search and cross-validation methods are combined to optimize the key hyperparameters of the random forest, constructing a dual-channel independent random forest regression prediction model with optimal generalization performance. Through the aforementioned periodic embedding and sub-objective prediction strategies, this application effectively overcomes the problem of discontinuous predictions at periodic boundaries in traditional random forest models, improves the extrapolation prediction capability of the dual-channel independent random forest regression prediction model for gust flow fields under unknown conditions, and significantly reduces the computational complexity associated with direct prediction of high-dimensional flow fields.

[0096] In the flow field reconstruction stage, this embodiment further constructs a gust waveform enhancement network based on the decoder structure. By establishing a mapping relationship between latent low-dimensional features and measurement point waveform information, waveform prediction constraints are introduced during training, enabling the gust flow field prediction model to simultaneously learn the evolution law of gust waveforms while reconstructing the velocity field. This mechanism, by introducing a waveform reconstruction error term (i.e., auxiliary loss) into the loss function, allows latent low-dimensional features to more fully reflect the physical information related to gust propagation, thereby effectively improving the physical relevance of latent variables and the overall interpretability of the gust flow field prediction model. Compared with traditional machine learning models that rely solely on data fitting, this embodiment effectively alleviates the "black box" problem of the gust flow field prediction model by adding physically relevant auxiliary constraints, improving the consistency and stability of the prediction results under different control parameter conditions.

[0097] In the training phase of the gust flow field prediction model, this application proposes an Auto-Stage Training method. By setting an automatic stage switching mechanism, the gust flow field prediction model prioritizes learning the main spatial structure of the flow field in the early stage of training, and gradually introduces waveform auxiliary constraints (i.e., auxiliary loss) in the later stage of training, achieving a smooth transition from single-objective optimization to multi-objective joint optimization. Simultaneously, this application employs an exponentially increasing weight function to dynamically adjust the auxiliary loss, and through a stage-related exponential growth time constant adjustment mechanism, automatically adjusts the weight growth rate according to the stage start time, ensuring that the gust flow field prediction model maintains good convergence characteristics in different training stages. Through the above dynamic training strategy, this application effectively avoids the training instability and weight selection difficulties that are prone to occur in traditional fixed-weight methods, improving the convergence speed and training stability of the gust flow field prediction model, thereby significantly improving the overall prediction accuracy.

[0098] The gust flow field prediction method of this application consists of a multi-source feature extraction module, a low-dimensional feature prediction module, a flow field reconstruction module, and a dynamic training and optimization module. These modules can either work collaboratively to form a complete prediction process or be called independently according to actual application needs, demonstrating strong flexibility. Furthermore, this method is not only applicable to the prediction of gust velocity fields but can also be extended to the evolution prediction of various flow parameters such as temperature, pressure, and vorticity fields. It can also be applied to different types of unsteady flow problems, thus significantly expanding the method's applicability and engineering application value.

[0099] Based on the same inventive concept, this application also provides a gust flow field prediction system for implementing the aforementioned gust flow field prediction method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more gust flow field prediction system embodiments provided below can be found in the limitations of the gust flow field prediction method described above, and will not be repeated here.

[0100] In one exemplary embodiment, such as Figure 6 As shown, a gust flow field prediction system is provided, comprising: The data acquisition and processing module is used to acquire gust flow field data and gust generator control parameters under different working conditions, and to construct training datasets and test datasets; the gust flow field data includes raw gust velocity field data and corresponding measurement point waveform information; The model building and training module is used to build a gust flow field prediction model and train the gust flow field prediction model based on the training dataset to obtain a trained gust flow field prediction model. The gust flow field prediction model includes an enhanced autoencoder model that integrates measurement point waveform information. The enhanced autoencoder model that integrates measurement point waveform information adopts an automatic stage switching mechanism and dynamic weights for staged training. It integrates measurement point waveform information including measurement point waveform velocity information and spatial position information on the basis of the autoencoder. By constructing a position-aware attention fusion mechanism, it achieves adaptive fusion of latent features and measurement point waveform features. The testing module is used to input the test dataset into the trained gust flow field prediction model to obtain gust velocity field prediction data. The prediction and evaluation module is used to compare and evaluate the predicted gust velocity field data with the original gust velocity field data in the test dataset to obtain the evaluation result.

[0101] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 7 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores gust flow field data under different operating conditions and gust generator control parameters. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a gust flow field prediction method.

[0102] Figure 7 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specifically, the computer device may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0103] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0104] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0105] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0106] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0107] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting gust flow fields, characterized in that, include: Acquire gust flow field data and gust generator control parameters under different working conditions, and construct training and testing datasets; The gust flow field data includes raw gust velocity field data and corresponding measurement point waveform information; A gust flow field prediction model is constructed and trained based on the training dataset to obtain a trained gust flow field prediction model. The gust flow field prediction model includes an enhanced autoencoder model that integrates measurement point waveform information. The enhanced autoencoder model that integrates measurement point waveform information adopts an automatic stage switching mechanism and dynamic weights for staged training. Measurement point waveform information, including measurement point waveform velocity information and spatial position information, is integrated on the basis of the autoencoder. By constructing a position-aware attention fusion mechanism, the adaptive fusion of latent features and measurement point waveform features is achieved. The enhanced autoencoder model that integrates waveform information from measurement points includes an encoding module and a decoding module; The encoding module is used to determine the fused potential low-dimensional features based on the training dataset; The automatic stage switching mechanism automatically switches between the first training stage and the second training stage based on the preset ratio between the flow field reconstruction loss and the initial flow field reconstruction loss. In the first training phase, the enhanced autoencoder model that integrates waveform information from measurement points takes flow field reconstruction error as the optimization objective, and the encoding module prioritizes learning the dominant spatial structure of the gust flow field to obtain the flow field reconstruction loss. In the second training phase, the enhanced autoencoder model that integrates the waveform information of the measurement points introduces a gust response prediction branch to determine the auxiliary loss, and uses an exponentially increasing weight function to perform weighted calculation on the flow field reconstruction loss and the auxiliary loss to determine the improved total loss; the gust response prediction branch is a branch of the decoding module in the enhanced autoencoder model that integrates the waveform information of the measurement points. The gust response prediction branch is used to calculate the velocity waveform prediction value based on the fused potential low-dimensional features, and the auxiliary loss represents the error loss between the velocity waveform prediction value and the measurement point waveform information. The test dataset is input into the trained gust flow field prediction model to obtain gust velocity field prediction data. The predicted gust velocity field data is compared and evaluated with the original gust velocity field data in the test dataset to obtain the evaluation results.

2. The method for predicting gust flow field according to claim 1, characterized in that, The encoding module is used to determine the fused potential low-dimensional features based on the training dataset, specifically including: The raw data of the gust velocity field and the waveform information of the measuring point are input into the encoding module of the enhanced autoencoder model that fuses the waveform information of the measuring point to obtain the latent features and the waveform features of the measuring point. A position-aware attention fusion mechanism is used to perform weighted fusion of the latent features and the waveform features of the measurement points to obtain the fused latent low-dimensional features.

3. The method for predicting gust flow field according to claim 2, characterized in that, The expression for the fused latent low-dimensional features is: ； in, This represents the potential low-dimensional features after fusion; Indicates potential characteristics; Indicates the waveform characteristics of the measuring point; For attention weights.

4. The method for predicting gust flow field according to claim 1, characterized in that, The automatic stage switching mechanism automatically switches between the first training stage and the second training stage based on a preset ratio between the flow field reconstruction loss and the initial flow field reconstruction loss. Specifically, it includes: When the flow field reconstruction loss is greater than a preset ratio of the initial flow field reconstruction loss, the enhanced autoencoder model that fuses the waveform information of the measurement points enters the first training stage, and the training loss of the enhanced autoencoder model that fuses the waveform information of the measurement points is the flow field reconstruction loss. When the flow field reconstruction loss is less than or equal to a preset proportion of the initial flow field reconstruction loss, the enhanced autoencoder model that fuses the waveform information of the measurement points automatically switches from the first training stage to the second training stage, and the training loss of the enhanced autoencoder model that fuses the waveform information of the measurement points is the improved total loss.

5. The method for predicting gust flow field according to claim 1, characterized in that, The expression for the switching condition of the automatic phase switching mechanism is: ； in, This represents the initial flow field reconstruction loss. This is the threshold coefficient for phase switching; This represents the loss due to flow field reconstruction. The expression for the improved total loss is as follows: ； in, The total loss after the improvement; It is an exponentially increasing weighting function; To mitigate losses; The expression for the exponentially increasing weighting function is: ； in, The maximum weight; This refers to the number of training rounds after entering the second training phase. It is the time constant for exponential growth.

6. The method for predicting gust flow field according to claim 5, characterized in that, The expression for the exponential growth time constant is: ； in, This represents the total number of training rounds. This represents the upper limit of the time constant. This is the lower limit of the time constant.

7. The method for predicting gust flow field according to claim 1, characterized in that, The gust flow field prediction model also includes a dual-channel independent random forest regression prediction model; The dual-channel independent random forest regression prediction model is used to determine the predicted potential low-dimensional features based on the training dataset and the fused potential low-dimensional features, specifically including: The dimensionless time parameter in the original gust velocity field data is calculated to obtain the periodic embedding feature; Based on the periodic embedding features and the gust generator control parameters, an input feature vector is constructed; The input feature vector is input into the dual-channel independent random forest regression prediction model, and the fused latent low-dimensional features are used as the output target to obtain the predicted latent low-dimensional features.

8. The method for predicting gust flow field according to claim 7, characterized in that, The dual-channel independent random forest regression prediction model uses mean squared error as the loss function during training. The expression for the loss function is: ； in, For hyperparameter combinations; For hyperparameter grid space A certain hyperparameter combination The calculated loss function; Indicates the first i The first actual latent low-dimensional feature of each sample Indicates the first i The first predicted latent low-dimensional feature of each sample; N The number of samples; i For sample index.

9. A gust flow field prediction system, characterized in that, The gust flow field prediction system is used to implement the gust flow field prediction method according to any one of claims 1-8, and the gust flow field prediction system includes: The data acquisition and processing module is used to acquire gust flow field data and gust generator control parameters under different working conditions, and to construct training datasets and test datasets; the gust flow field data includes raw gust velocity field data and corresponding measurement point waveform information; The model building and training module is used to build a gust flow field prediction model and train the gust flow field prediction model based on the training dataset to obtain a trained gust flow field prediction model. The gust flow field prediction model includes an enhanced autoencoder model that integrates measurement point waveform information. The enhanced autoencoder model that integrates measurement point waveform information adopts an automatic stage switching mechanism and dynamic weights for staged training. It integrates measurement point waveform information including measurement point waveform velocity information and spatial position information on the basis of the autoencoder. By constructing a position-aware attention fusion mechanism, it achieves adaptive fusion of latent features and measurement point waveform features. The enhanced autoencoder model that integrates waveform information from measurement points includes an encoding module and a decoding module; The encoding module is used to determine the fused potential low-dimensional features based on the training dataset; The automatic stage switching mechanism automatically switches between the first training stage and the second training stage based on the preset ratio between the flow field reconstruction loss and the initial flow field reconstruction loss. In the first training phase, the enhanced autoencoder model that integrates waveform information from measurement points takes flow field reconstruction error as the optimization objective, and the encoding module prioritizes learning the dominant spatial structure of the gust flow field to obtain the flow field reconstruction loss. In the second training phase, the enhanced autoencoder model that integrates the waveform information from the measurement points introduces a gust response prediction branch to determine the auxiliary loss. An exponentially increasing weighting function is used to weight the flow field reconstruction loss and the auxiliary loss to determine the improved total loss. The gust response prediction branch is a branch of the decoding module in the enhanced autoencoder model that integrates the waveform information from the measurement points. The gust response prediction branch is used to calculate the predicted velocity waveform value based on the fused latent low-dimensional features. The auxiliary loss represents the error loss between the predicted velocity waveform value and the waveform information from the measurement points. The testing module is used to input the test dataset into the trained gust flow field prediction model to obtain gust velocity field prediction data. The prediction and evaluation module is used to compare and evaluate the predicted gust velocity field data with the original gust velocity field data in the test dataset to obtain the evaluation result.

Images