Wind speed three-dimensional profile intelligent reconstruction method and system based on deep learning

Abstract

The invention belongs to the technical field of wind profiles, and discloses a wind speed three-dimensional profile intelligent reconstruction method and system based on deep learning. The method comprises the following steps: obtaining regional wind field data of a sea area in a certain period of time, performing two-dimensional slicing in a spatial dimension to serve as input data of a teacher model, then building the teacher model and training, and outputting an intermediate feature map in the middle of the teacher model; adopting intermediate feature distillation to align intermediate feature maps output by the teacher model and the student model on respective networks, and calculating an error between the intermediate feature map output by the teacher network and the intermediate feature map output by the student network; and building a student model and training, and deducing a complete three-dimensional wind profile of the certain position point through the student model. According to the method, a small-parameter network structure is realized, the operation burden is reduced, the student model only needs to process single-point input during reasoning, the prediction speed is high, and the requirement of actual engineering for quick response can be met.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind profile, and particularly relates to a wind speed three-dimensional profile intelligent reconstruction method and system based on deep learning. BACKGROUND

[0002] Wind profile refers to the vertical distribution structure of wind speed and wind direction changing with height, and is an important representation of atmospheric boundary layer dynamic process. The wind profile not only directly affects the accuracy of atmospheric dynamics research and numerical model, but also has key significance for application scenarios such as aviation flight safety, ship navigation stability, offshore engineering construction and wind energy development. For example, in the process of carrier-based aircraft taking off and landing, the vertical shear and turbulent structure of local three-dimensional wind profile directly determine the safety of the aircraft; in the process of wind farm site selection and operation, the wind speed distribution at different heights is an important basis for evaluating wind energy resources.

[0003] At present, the method for obtaining wind profile mainly depends on two ways: one is observation means, such as radiosonde, laser radar, etc., which can directly provide local vertical wind field information, but has the limitations of limited spatial coverage, high cost, and difficulty in deployment in complex ocean and mobile platform environment; the other is numerical mode calculation, such as numerical mode, which can simulate regional wind field profile, but has high calculation cost and is difficult to meet the real-time requirement. In recent years, some researches have tried to use deep learning model to infer vertical wind field from low-level wind speed, but most of them rely on large-scale ground wind field as input, which has high input data requirement and cannot directly obtain accurate three-dimensional wind profile under the condition of only single-point observation.

[0004] At present, an invention patent discloses a wind profile calculation method and system based on deep learning (publication number CN117034761A, publication date 20231110), which uses a deep learning model to learn the mapping relationship between temperature and wind direction and wind speed data at different heights. It uses global upper air meteorological observation data as a data set, which requires a large amount of observation data for model prediction.

[0005] An existing invention patent discloses a wind profile radar boundary layer height inversion method based on deep learning (publication number CN118915006A, publication date 20241108), which uses wind profile radar product data to predict the boundary layer height. However, the model adds various atmospheric parameter profiles, and the model based on empirical formula inevitably has some differences from the real situation.

[0006] The prior art (Lin C, Tie R, Yi S, et al. Terrain-aware Deep Learning for Wind Energy Applications: From Kilometer-scale Forecasts to Fine Wind Fields [J]. 2025.) proposes a three-dimensional wind reconstruction method based on deep learning, which uses a large amount of CFD simulation data and uses regional wind and ground elements as input to reconstruct three-dimensional wind field. It is difficult to obtain simulation data of this scale for prediction in reality and it is difficult to achieve real-time in application.

[0007] Through the above analysis, the problems and defects of the prior art are: the prior art needs to input complete regional ground wind field or multi-level observation data to obtain high-precision wind profile results, which is difficult to popularize and apply in the scene of limited actual observation data; in order to fit the complex spatial relationship of wind field, the network structure with large number of parameters is generally used, or the experience formula is used to interfere with the model learning, which leads to high consumption of computing resources, low model precision, and low wind speed prediction speed and precision, which is not conducive to deployment in application scenarios with high real-time requirements. SUMMARY

[0008] In order to overcome the problems in the related art, the present application discloses an intelligent reconstruction method for wind speed three-dimensional profile based on deep learning and a system, which specifically relates to an intelligent reconstruction method for vertical wind profile.

[0009] The technical solution is as follows: an intelligent reconstruction method for wind speed three-dimensional profile based on deep learning, comprising the following steps: S1, 10m wind speed and three-dimensional wind field data preparation: using WRF numerical model to simulate the wind speed component at 10m height of a sea area and the three-dimensional wind speed field in the vertical direction to obtain the regional wind field data of the sea area within a certain period of time as the data set of the subsequent model, taking as input, as output label; the regional wind field data includes the ground layer 10m wind component , ; the three-dimensional wind field component , , wherein is the vertical height level, is the time dimension value, is the horizontal longitude coordinate, is the vertical latitude coordinate; S2, Teacher Model Construction and Training: The 10m wind speed and three-dimensional wind speed data of this sea area are divided into two-dimensional slices in the spatial dimension and used as input data for the teacher model. The teacher model is built and trained, and intermediate feature maps are output in the middle of the teacher model. S3, Design Knowledge Distillation: Use intermediate feature distillation to align the intermediate feature maps output by the teacher model and the student model on their respective networks, and calculate the error between the intermediate feature maps output by the teacher network and the intermediate feature maps output by the student network. S4, Student Model Construction and Training: The 10m wind speed and 3D wind speed data of this sea area are spatially sliced ​​into two-dimensional single-point data points. These are used as input data for the student model. The student model is built and trained, using supervised loss based on error and true values, and distillation loss based on intermediate features with the teacher model. Total loss will result; S5, Model Application Based on Single-Point Input: After both the teacher and student models have been trained, in the practical application stage, inputting a specific location and time... The complete three-dimensional wind profile at that location is inferred using the student model.

[0010] In step S2, the 10m wind speed and three-dimensional wind speed data of the sea area are sliced ​​in two dimensions in the spatial dimension, including: Based on window size At the same time, on the same horizontal grid, for and the three-dimensional reference height layers selected. The wind component is sliced ​​and concatenated according to channels to form the input tensor, expressed as: ; ; In the formula, Index of the window's starting position. For this position at this time Input slice, For the vertical three-dimensional position at this time Wind speed slice, To use step size as The range of vertical coordinates, Let the side length of the window be 1. To use step size as The range of horizontal coordinates.

[0011] Furthermore, a teacher model is built and trained, and intermediate feature maps are output from the teacher model, including: The teacher network used to build the teacher model adopts a convolutional neural network structure of input layer-feature extraction layer-output layer. After the input layer receives the tensor X of each batch, it enters the feature extraction layer, which consists of four concatenated convolutional blocks. Each convolutional block contains 3×3 convolution, batch normalization (BN), and ReLU activation. The convolution stride is 1 and the boundary padding is 1 to keep the feature map size consistent during the extraction process. After the second convolutional block, a 2×2 max pooling is performed. The number of channels in the network is gradually increased from shallow to deep. Then, at the bottleneck, the feature map is flattened into a vector and fed into a two-layer fully connected regression head to regress the wind profile of the center grid point of the slice on the height layer in one go. Several inputs are used during teacher model training. , , , This allows the teacher model to learn the mapping relationship between 10m wind slices and 3D wind slices, through fixed time intervals. and fixed position The corresponding wind profile is obtained, and its expression is: ; In the formula, This represents the vertical wind profile at the corresponding moment. Input for the model Quantity, Input for the model Quantity, The three-dimensional V-direction wind component is the input to the model. The wind profile output for the teacher model; The teacher network selects one intermediate feature map from the shallow layer and one from the deep layer as cue signals, and derives them at the output of the second convolutional block. The feature map is derived at the output of the fourth convolutional block. Feature map; Feature map The feature maps correspond to the intermediate and final features of the teacher model, respectively, and are indexed by sample. Archive caching; the training objective of the teacher model is to minimize the mean squared error loss between the predicted profile and the true profile: ; In the formula, The vertical wind profile reconstructed for the teacher model. For the true wind profile in the tag set, This represents the total number of samples output by the model. The optimizer uses Adam, and after the validation set converges, all parameters of the teacher network and intermediate features are frozen and derived.

[0012] In step S3, intermediate feature maps of the teacher model and student model outputs on their respective networks are aligned using intermediate feature distillation, including: After the teacher network training converges and parameters are frozen, intermediate feature distillation is used to align features of the teacher model and the student model at different semantic levels; the outputs of the second and fourth convolutional blocks of the teacher model are fixed as two levels of intermediate features. and For the same spatiotemporal index in the same batch For the given samples, the teacher model produces two-level feature maps and a target profile, while the student model obtains two-level student features of the same size as the teacher model in its second hidden layer. During training, the student model simultaneously minimizes two parts of the loss: the supervised loss and the loss based on the true values. Distillation loss of intermediate features with the teacher model During distillation, the teacher and student models are based on the same sample index. Comparison of intermediate features and ,in, Representative teacher model, Representative student model.

[0013] In step S4, the student network for building the student model adopts a three-layer fully connected structure of input layer-topic layer-output layer. The main body of the network is a three-layer fully connected regression backbone MLP, and the input is a single point. The two-dimensional vector is passed through two hidden layers and one output layer: the first hidden layer is a linear mapping of channels 2 to 64 followed by ReLU activation; the second hidden layer is a linear mapping of channels 64 to 128 followed by ReLU activation; the output layer is a linear mapping of channels 128 to 20, based on the wind profile regression results at a height layer of 10; the student model uses a single-point 10m wind speed component. The output is the vertical 3D wind profile of that point, expressed as: ; In the formula, , This represents the vertical wind profile at the corresponding moment. The mapping function for the student model uses the single-point 10m wind speed component as input. The u-direction component of the vertical three-dimensional wind profile obtained from the reconstruction of the student model. The v-direction component of the vertical three-dimensional wind profile obtained by reconstructing the student model.

[0014] Furthermore, the total loss of the student model is: ; In the formula, The supervised loss is the true value. Distillation loss of intermediate features in the teacher model. , The adjustable weighting coefficients represent the supervision loss relative to the true value. Distillation loss of intermediate features with the teacher model The weighting of the two parts of the loss in the total loss is calculated as follows: ; ; In the formula, , These represent the feature maps of the teacher model and the student model at the same data index, respectively. For the true wind profile in the tag set, The horizontal wind speed u component at each height in the vertical direction predicted by the student model. The horizontal wind speed v component at each height in the vertical direction is predicted by the student model.

[0015] Another objective of this invention is to provide a deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles. This system implements the deep learning-based intelligent reconstruction method for three-dimensional wind speed profiles and includes: The 10m wind speed and 3D wind field data preparation module is used to prepare wind speed components at a height of 10m in a certain sea area using the WRF numerical model. and the three-dimensional wind speed field in the vertical direction Numerical simulations were conducted to obtain regional wind field data for a certain period of time in the sea area, which was then used as the dataset for subsequent models. As input, As output labels; regional wind field data includes the wind component at 10m above ground level. , ; Three-dimensional wind field components , ; The teacher model building and training module is used to slice the 10m wind speed and three-dimensional wind speed data of the sea area into two-dimensional slices in the spatial dimension, which are used as input data for the teacher model, to build and train the teacher model, and to output intermediate feature maps in the middle of the teacher model. Design a knowledge distillation process module to align the intermediate feature maps output by the teacher model and the student model on their respective networks using intermediate feature distillation, and calculate the error between the intermediate feature maps output by the teacher network and the student network. The student model building and training module is used to slice the 10m wind speed and 3D wind speed data of this sea area into two-dimensional spatial single-point data in a spatial dimension. This data is used as the input data for the student model to build and train the student model. The module uses supervised loss based on error and true values, and distillation loss based on intermediate features with the teacher model. Total loss will result; The model application module based on single-point input is used in the practical application stage after both the teacher model and the student model have been trained, by inputting a specific location and time. The complete three-dimensional wind profile at that location is inferred using the student model.

[0016] Furthermore, the system is mounted on a computer device, which includes at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, it implements the functions of the aforementioned deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles.

[0017] Furthermore, the system is mounted on a computer-readable storage medium that stores a computer program. When the computer program is executed by a processor, it can realize the functions of the aforementioned deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles.

[0018] Furthermore, the system is mounted on an information data processing terminal, which, when executed on an electronic device, provides a user input interface to implement the functions described above in the deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles. Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention applies a teacher-student knowledge distillation model. Employing a knowledge distillation mechanism, during the training phase, it not only utilizes real 3D wind profiles as supervision but also, through feature layer distillation, enables the student model to directly learn the intermediate feature representations of the teacher model. This method effectively shortens the convergence time of the student model while maintaining predictive ability close to that of the teacher model. In practical applications, this invention allows student models to reconstruct a complete 3D wind profile with only a single-point 10m wind speed input, eliminating the need for regionalized input. This design significantly reduces input requirements and computational overhead, offering advantages such as lightweight design, real-time performance, and ease of engineering deployment.

[0019] This invention employs a teacher-student knowledge distillation framework: the teacher model establishes a mapping relationship between 10m wind speeds over a large area and 3D wind profiles, while the student model learns under the teacher's guidance. During the testing phase, it only needs to input a single 10m wind speed to intelligently reconstruct the complete 3D wind profile at that location. Compared with existing technologies, it has the following significant advantages: Low input requirements: It eliminates the dependence on large-area wind field or multi-layer observation data, requiring only a single 10m wind speed to complete the 3D profile inference. Lightweight and real-time: Through knowledge distillation, the complex knowledge of the teacher model is transferred to the student model, achieving a network structure with a small number of parameters, reducing the computational burden. Furthermore, the student model only needs to process a single-point input during inference, resulting in fast prediction speed, which can meet the needs of practical engineering for rapid response.

[0020] Secondly, this invention can significantly reduce the amount of data and computing resources required for wind profile reconstruction, providing a lightweight and real-time intelligent wind field reconstruction method for fields such as wind energy development, ship navigation safety assurance, and aviation meteorological forecasting. Since vertical profile prediction can be achieved with only a single-point wind speed of 10m, this method can be quickly embedded into observation or operation platforms such as offshore buoys, unmanned vessels, and drones, possessing advantages of high versatility and low-cost deployment.

[0021] Current research on deep learning wind profiles, both domestically and internationally, largely relies on regional grid wind fields or multi-layer observation data. These models are complex and lack real-time performance, making them difficult to deploy in resource-constrained environments. This invention, for the first time, proposes introducing a "teacher-student knowledge distillation" mechanism into the task of reconstructing three-dimensional wind speed profiles. This enables the mapping learning from single-point ground wind speeds to complete vertical profiles, overcoming the limitation of traditional methods that require large-scale wind field input.

[0022] For a long time, wind profile inversion has been limited by sparse vertical layer observations, insufficient resolution of numerical simulations, and huge computational costs, making it difficult to achieve real-time, fine-grained wind field reconstruction. This invention introduces a knowledge distillation framework to transfer the high-dimensional wind field knowledge of a complex teacher model to a lightweight student model. This allows the student model to accurately infer the three-dimensional wind speed distribution even with only a single-point wind speed input, effectively solving the technical challenge of "accurate wind profile recovery under low input conditions" and achieving a balance between model lightweighting and prediction accuracy.

[0023] Traditionally, it is widely believed that wind profile reconstruction requires multidimensional wind field input or dense observation data; otherwise, high-precision results cannot be obtained. This invention, through systematic experimental verification, demonstrates that even a single-point 10m wind speed input can learn the distribution patterns of high-dimensional wind field characteristics within a teacher-student distillation framework. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the intelligent reconstruction method for three-dimensional wind speed profile based on deep learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the three-dimensional U-shaped profile at a certain position at a certain moment, provided by the present invention. Figure 3 This is a schematic diagram of the three-dimensional V-profile at a certain position at a certain moment, provided by the present invention. Figure 4 This is a comparison diagram of the three-dimensional wind field components U of the reconstructed wind profile provided by this invention and other methods; Figure 5 This is a comparison diagram of the three-dimensional wind field components V of the reconstructed wind profile provided by the present invention and other methods. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0026] The innovation of this invention lies in its proposal of an intelligent reconstruction method for 3D wind speed profiles based on a teacher-student knowledge distillation framework. This method learns high-dimensional wind speed distribution patterns in a large-scale wind field using a teacher model, and then transfers this knowledge to a simplified student model. This enables high-precision 3D profile reconstruction with only a single-point 10m wind speed input. This method fully leverages the advantages of parameter compression and feature transfer from knowledge distillation, significantly reducing the number of model parameters and computational load, and substantially improving the model's lightweight nature and inference speed. It maintains prediction accuracy while meeting the real-time performance and low resource consumption requirements of practical engineering applications, demonstrating good deployability and application value.

[0027] Example 1, such as Figure 1 As shown, the intelligent reconstruction method for three-dimensional wind speed profile based on deep learning provided in this embodiment of the invention includes: Preparation of S1, 10m wind speed and 3D wind field data: The wind speed components at a height of 10m in a certain sea area were analyzed using the WRF numerical model. and the three-dimensional wind speed field in the vertical direction Numerical simulations were conducted to obtain regional wind field data for a certain period of time in the sea area, which was then used as the dataset for subsequent models. As input, As output labels; regional wind field data includes the wind component at 10m above ground level. , ; Three-dimensional wind field components , ,in, Vertical height hierarchy, Values ​​are in the time dimension. The coordinates are lateral longitude coordinates. Vertical latitude coordinates; For example, acquire regional 10m wind field data and three-dimensional wind field data, and organize these data into the size and shape of the model input.

[0028] Specifically, wind field data for the study area are obtained from numerical models or observational data, including the 10m surface wind component. , ; S2, Teacher Model Construction and Training: The 10m wind speed and three-dimensional wind speed data of this sea area are divided into two-dimensional slices in the spatial dimension and used as input data for the teacher model. The teacher model is built and trained, and intermediate feature maps are output in the middle of the teacher model. Specifically, first by window size (e.g., k=9) At the same time, on the same horizontal grid, for and the three-dimensional reference height layers selected. The wind component is sliced ​​and stitched together according to the channels to form the input tensor: ; ; In the formula, Index of the window's starting position. For this position at this time Input slice, For the vertical three-dimensional position at this time Wind speed slice, To use step size as The range of vertical coordinates, Let the side length of the window be 1. To use step size as The range of horizontal coordinates.

[0029] Furthermore, the teacher network adopts a convolutional neural network structure of "input layer - feature extraction layer - output layer": the input layer receives the tensor X of each batch and then enters the feature extraction layer, which consists of four concatenated convolutional blocks. Each convolutional block contains 3×3 convolutions, batch normalization (BN), and ReLU activation, with a convolution stride of 1 and boundary padding of 1 to keep the feature map size consistent during extraction; a 2×2 max pooling is performed after the second convolutional block. The number of channels in the network gradually increases from shallow to deep. Then, at the bottleneck, the feature map is flattened into a vector and fed into a two-layer fully connected regression head to regress the wind profile of the center grid point of the slice on the height layer in one go. That is, the teacher model is trained by inputting the above-mentioned tensors X. , , , This allows the teacher model to learn the mapping relationship between 10m wind slices and 3D wind slices, through fixed time intervals. Fixed position The corresponding wind profile can be obtained, namely: ; In the formula, This represents the vertical wind profile at the corresponding moment. Input for the model Quantity, Input for the model Quantity, The three-dimensional V-direction components are input to the model. The wind profile output for the teacher model; To ensure a smooth transition to the subsequent distillation process, the teacher network selects a representative intermediate feature map from both the shallow and deep layers as a "cue signal," which is then exported at the output of the second convolutional block. The feature map is derived at the output of the fourth convolutional block. Feature maps. These correspond to the intermediate and final features of the teacher model, respectively, and are indexed by sample. Archive caching. The training objective on the teacher's side is to minimize the mean squared error loss between the predicted profile and the true profile: ; In the formula, The vertical wind profile reconstructed for the teacher model. For the true wind profile in the tag set, This represents the total number of samples output by the model. Furthermore, the optimizer uses Adam (with an initial learning rate of 0.001 and cosine annealing), and after the validation set converges, all parameters of the teacher network and intermediate features are frozen and derived.

[0030] For example, in this invention, a regionalized 10m wind speed field is used as input in the teacher model, and a deep convolutional neural network is used to learn the mapping relationship between the input and the three-dimensional wind profile. The teacher model can comprehensively capture large-scale spatial information and output a complete three-dimensional wind profile. Simultaneously, the teacher model also generates intermediate feature representations during training. These features contain the mapping relationship between the input and output, providing guidance for the learning of the student model.

[0031] S3, Design Knowledge Distillation: Use intermediate feature distillation to align the intermediate feature maps output by the teacher model and the student model on their respective networks, and calculate the error between the intermediate feature maps output by the teacher network and the intermediate feature maps output by the student network. The MSE of both was calculated. Therefore, the loss of the student model consists of two parts: the supervision loss relative to the true value. Distillation loss of intermediate features with the teacher model .

[0032] Specifically, after the teacher network training converges and parameters are frozen, intermediate feature distillation is used to align the intermediate feature maps output by the teacher and student models after the convolutional layers. These intermediate feature maps reflect the spatial distribution patterns and gradient changes of the wind speed learned by the two models. The outputs of the second and fourth convolutional blocks of the teacher model are fixed as two levels of intermediate features. and For the same spatiotemporal index in the same batch The teacher model outputs two-level feature maps and a target profile from the sample data, while the student model outputs two-level intermediate feature maps of the same size as the teacher model in its hidden layer. Thus, the student model minimizes two parts of the loss during training: the supervised loss and the loss from the true values. Distillation loss of intermediate features with the teacher model During distillation, the teacher model and the student model must be based on the same sample index. Comparison of intermediate features and ,in, Representative teacher model, Representative student model.

[0033] For example, to transfer the complex knowledge contained in the teacher model to a lightweight student model, this invention designs a distillation mechanism. The student model not only learns realistic 3D wind profile labels but also mimics the intermediate layer feature distribution of the teacher model through feature matching. By introducing a distillation loss function, the student model can converge quickly under the teacher's guidance and achieve expressive power close to that of the teacher model.

[0034] S4, Student Model Construction and Training: The 10m wind speed and 3D wind speed data of this sea area are spatially sliced ​​into two-dimensional single-point data points. These are used as input data for the student model. The student model is built and trained, using supervised loss based on error and true values, and distillation loss based on intermediate features with the teacher model. Total loss will result; Furthermore, the student network adopts a three-layer fully connected structure of "input layer-topic layer-output layer," with the main body of the network being a three-layer fully connected regression backbone (MLP). The input is a single point. The two-dimensional vector is passed through two hidden layers and one output layer: the first hidden layer is a linear mapping of channels 2 to 64 followed by ReLU activation; the second hidden layer is a linear mapping of channels 64 to 128 followed by ReLU activation; the output layer is a linear mapping of channels 128 to 20, based on the wind profile regression result at height layer 10. That is, the student model uses a single-point 10m wind speed component. The output is the vertical 3D wind profile of that point, expressed as: ; In the formula, , This represents the vertical wind profile at the corresponding moment. The mapping function for the student model uses the single-point 10m wind speed component as input. The u-direction component of the vertical three-dimensional wind profile obtained from the reconstruction of the student model. The v-direction component of the vertical three-dimensional wind profile obtained by reconstructing the student model.

[0035] The total loss of the student model is: ; In the formula, The supervised loss is the true value. Distillation loss of intermediate features in the teacher model. , The adjustable weighting coefficients represent the supervision loss relative to the true value. Distillation loss of intermediate features with the teacher model The weighting of the two parts of the loss in the total loss is calculated as follows: ; ; In the formula, , These represent the feature maps of the teacher model and the student model at the same data index, respectively. For the true wind profile in the tag set, The horizontal wind speed u component at each height in the vertical direction predicted by the student model. The horizontal wind speed v component at each height in the vertical direction is predicted by the student model.

[0036] In this way, the student model can retain the feature representation ability of the teacher model even with limited input information.

[0037] For example, a student model is built, with the student model only focusing on a single point. As input, the data passes through a lightweight network, which outputs a 3D wind speed profile for that point. Because the student model receives guidance from the teacher model during training, it is able to make inferences with limited input information. This significantly reduces the model's input requirements and computational complexity.

[0038] S5, Model Application Based on Single-Point Input: After both the teacher and student models have been trained, in the practical application stage, inputting a specific location and time... The complete three-dimensional wind profile at that location is inferred using the student model.

[0039] For a test dataset consisting of a single-point 10m wind speed, a rapid reconstruction of the corresponding three-dimensional wind profile is achieved.

[0040] For example, after the entire teacher-student knowledge distillation model has been trained, only a specific time and position needs to be input into the student model test. This allows you to quickly output the complete vertical wind profile of that point.

[0041] Furthermore, this invention uses the PyTorch framework for model building, training, and testing.

[0042] Further, export the pth weight file after training is complete.

[0043] This invention uses a teacher-student knowledge distillation model to capture features of the mapping relationship between 10m wind data and 3D wind data. The teacher model learns over a large area. With three dimensions The mapping relationship between them, the student model learns the single point perpendicular to it The mapping relationship between them is established, and intermediate features of the teacher model are learned. After all training is complete, the student model only needs to input a single point during the inference phase. The wind speed alone can intelligently reconstruct the complete three-dimensional wind profile at a given location. This invention eliminates the reliance of most current models on large-scale regional wind fields or multi-layer observation data during the inference stage, requiring only a single-point wind speed of 10m to complete the three-dimensional profile inference; furthermore, through knowledge distillation, the complex knowledge of the teacher model is transferred to the student model, achieving a lightweight network structure and parameters.

[0044] Example 2: This invention provides a deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles, comprising: The 10m wind speed and 3D wind field data preparation module is used to prepare wind speed components at a height of 10m in a certain sea area using the WRF numerical model. and the three-dimensional wind speed field in the vertical direction Numerical simulations were conducted to obtain regional wind field data for a certain period of time in the sea area, which was then used as the dataset for subsequent models. As input, As output labels; regional wind field data includes the wind component at 10m above ground level. , ; Three-dimensional wind field components , ; The teacher model building and training module is used to slice the 10m wind speed and three-dimensional wind speed data of the sea area into two-dimensional slices in the spatial dimension, which are used as input data for the teacher model, to build and train the teacher model, and to output intermediate feature maps in the middle of the teacher model. Design a knowledge distillation process module to align the intermediate feature maps output by the teacher model and the student model on their respective networks using intermediate feature distillation, and calculate the error between the intermediate feature maps output by the teacher network and the student network. The student model building and training module is used to slice the 10m wind speed and 3D wind speed data of this sea area into two-dimensional spatial single-point data in a spatial dimension. This data is used as the input data for the student model to build and train the student model. The module uses supervised loss based on error and true values, and distillation loss based on intermediate features with the teacher model. Total loss will result; The model application module based on single-point input is used in the practical application stage after both the teacher model and the student model have been trained, by inputting a specific location and time. The complete three-dimensional wind profile at that location is inferred using the student model.

[0045] To further illustrate the effects of the embodiments of the present invention, the following experiments were conducted.

[0046] Based on the above steps, the WRF numerical model was used to obtain the following data for the Chinese sea area shown in the figure below, following step S1. Data and 3D data, The data size is 4224×216×216, and the data dimensions represent time, spatial y-axis and spatial x-axis respectively; The data size is 4224×10×216×216, and the data dimensions represent time, spatial height, spatial y-axis and spatial x-axis respectively.

[0047] Following step S2, the acquired data is sliced ​​using a 9×9 spatial window and a reverse slice method to obtain... Input fragment and its corresponding 3D Sample. Each The input sample size is (2, 9, 9), three-dimensional The sample size is (10, 1, 1). The samples are fed into the model in batches according to the batch size for training. A teacher model is built using a CNN model (4×Conv, 1×MaxPool, Flatten, FC). The dataset is divided into training set, validation set and test set in a ratio of 3:1:1.

[0048] Follow step S3 to set the supervised loss relative to the true value. The weighting is 0.7 and the distillation loss is based on the intermediate features of the teacher model. The weighting is 0.3.

[0049] Follow step S4 to build a lightweight student model, which is a lightweight MLP model, and each training data input... The sample size is (2, 1, 1), and it is three-dimensional. The sample size is (10, 1, 1), and the samples are fed into the model in batches according to the batch size for training.

[0050] Following step S5, after all models have been trained, the 10m wind speed value at a certain location at a certain time is used. For testing purposes, you can quickly obtain the vertical 10 floors of the store. Profiles require no regional wind field input, resulting in high computational efficiency. For example... Figure 2 The model outputs a schematic diagram of the 3D U-shaped profile at a specific time and location; for example... Figure 3 The model outputs a schematic diagram of the three-dimensional V-profile at a certain time and location.

[0051] Furthermore, for the test wind field data, the traditional method (the empirical power law formula of wind profile) and the student model without distillation mechanism were used as control groups. The accuracy of the model constructed in this invention was compared by calculating the average MAE, RMSE and R, as shown in Table 1. Table 1 Comparison of Average MAE, RMSE, and R

[0052] Compare the reconstructed profiles of different methods with the actual profiles at a specific location and time. Figure 4 This is a comparison diagram of the three-dimensional wind field components U of the reconstructed wind profile provided by this invention and other methods; Figure 5 This is a comparison diagram of the three-dimensional wind field components V of the reconstructed wind profile provided by the present invention and other methods.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent reconstruction of three-dimensional wind speed profiles based on deep learning, characterized in that, The method includes the following steps: Preparation of S1, 10m wind speed and 3D wind field data: The wind speed components at a height of 10m in a certain sea area were analyzed using the WRF numerical model. and the three-dimensional wind speed field in the vertical direction Numerical simulations were conducted to obtain regional wind field data for a certain period of time in the sea area, which was then used as the dataset for subsequent models. As input, As output labels; regional wind field data includes the wind component at 10m above ground level. , ; Three-dimensional wind field components , ,in, Vertical height hierarchy Values ​​are in the time dimension. The coordinates are lateral longitude coordinates. Vertical latitude coordinates; S2, Teacher Model Construction and Training: The 10m wind speed and three-dimensional wind speed data of this sea area are divided into two-dimensional slices in the spatial dimension and used as input data for the teacher model. The teacher model is built and trained, and intermediate feature maps are output in the middle of the teacher model. S3, Design Knowledge Distillation: Use intermediate feature distillation to align the intermediate feature maps output by the teacher model and the student model on their respective networks, and calculate the error between the intermediate feature maps output by the teacher network and the intermediate feature maps output by the student network. S4, Student Model Construction and Training: The 10m wind speed and 3D wind speed data of this sea area are spatially sliced ​​into two-dimensional single-point data points. These are used as input data for the student model. The student model is built and trained, using supervised loss based on error and true values, and distillation loss based on intermediate features with the teacher model. Total loss will result; S5, Model Application Based on Single-Point Input: After both the teacher and student models have been trained, in the practical application stage, inputting a specific location and time... The complete three-dimensional wind profile at that location is inferred using the student model.

2. The intelligent reconstruction method for three-dimensional wind speed profile based on deep learning according to claim 1, characterized in that, In step S2, the 10m wind speed and three-dimensional wind speed data of the sea area are sliced ​​in two dimensions in the spatial dimension, including: Based on window size At the same time, on the same horizontal grid, for and the three-dimensional reference height layers selected. The wind component is sliced ​​and concatenated according to channels to form the input tensor, expressed as: ； ； In the formula, Index of the window's starting position. For this position at this time Input slice, For the vertical three-dimensional position at this time Wind speed slice, To use step size as The range of vertical coordinates, Let the side length of the window be 1. To use step size as The range of horizontal coordinates.

3. The intelligent reconstruction method for three-dimensional wind speed profile based on deep learning according to claim 2, characterized in that, Build and train a teacher model, and output intermediate feature maps from the teacher model, including: The teacher network used to build the teacher model adopts a convolutional neural network structure of input layer-feature extraction layer-output layer. After the input layer receives the tensor X of each batch, it enters the feature extraction layer, which consists of four concatenated convolutional blocks. Each convolutional block contains 3×3 convolution, batch normalization (BN), and ReLU activation. The convolution stride is 1 and the boundary padding is 1 to keep the feature map size consistent during the extraction process. After the second convolutional block, a 2×2 max pooling is performed. The number of channels in the network is gradually increased from shallow to deep. Then, at the bottleneck, the feature map is flattened into a vector and fed into a two-layer fully connected regression head to regress the wind profile of the center grid point of the slice on the height layer in one go. Several inputs are used during teacher model training. , , , This allows the teacher model to learn the mapping relationship between 10m wind slices and 3D wind slices, through fixed time intervals. and fixed position The corresponding wind profile is obtained, and its expression is: ； In the formula, This represents the vertical wind profile at the corresponding moment. Input for the model Quantity, Input for the model Quantity, 3D input for the model Directional wind component, The wind profile output for the teacher model; The teacher network selects one intermediate feature map from the shallow layer and one from the deep layer as cue signals, and derives them at the output of the second convolutional block. The feature map is derived at the output of the fourth convolutional block. Feature map; Feature map The feature maps correspond to the intermediate and final features of the teacher model, respectively, and are indexed by sample. Archive caching; the training objective of the teacher model is to minimize the mean squared error loss between the predicted profile and the true profile: ； In the formula, The vertical wind profile reconstructed for the teacher model. For the true wind profile in the tag set, This represents the total number of samples output by the model. The optimizer uses Adam, and after the validation set converges, all parameters of the teacher network and intermediate features are frozen and derived.

4. The intelligent reconstruction method for three-dimensional wind speed profile based on deep learning according to claim 1, characterized in that, In step S3, intermediate feature maps of the teacher model and student model outputs on their respective networks are aligned using intermediate feature distillation, including: After the teacher network training converges and parameters are frozen, intermediate feature distillation is used to align features of the teacher model and the student model at different semantic levels; the outputs of the second and fourth convolutional blocks of the teacher model are fixed as two levels of intermediate features. and For the same spatiotemporal index in the same batch For the given samples, the teacher model produces two-level feature maps and a target profile, while the student model obtains two-level student features of the same size as the teacher model in its second hidden layer. During training, the student model simultaneously minimizes two parts of the loss: the supervised loss and the loss based on the true values. Distillation loss of intermediate features with the teacher model During distillation, the teacher and student models are based on the same sample index. Comparison of intermediate features and ,in, Representative teacher model, Representative student model.

5. The intelligent reconstruction method for three-dimensional wind speed profile based on deep learning according to claim 1, characterized in that, In step S4, the student network for building the student model adopts a three-layer fully connected structure of input layer-topic layer-output layer. The main body of the network is a three-layer fully connected regression backbone MLP, and the input is a single point. The two-dimensional vector is passed through two hidden layers and one output layer: the first hidden layer is a linear mapping of channels 2 to 64 followed by ReLU activation; the second hidden layer is a linear mapping of channels 64 to 128 followed by ReLU activation; the output layer is a linear mapping of channels 128 to 20, based on the wind profile regression results at a height layer of 10; the student model uses a single-point 10m wind speed component. The output is the vertical 3D wind profile of that point, expressed as: ； In the formula, , This represents the vertical wind profile at the corresponding moment. The mapping function for the student model uses the single-point 10m wind speed component as input. The wind speed component in the U-direction of the vertical three-dimensional wind profile obtained from the reconstruction of the student model. The V-direction wind speed component of the vertical three-dimensional wind profile obtained from the reconstruction of the student model.

6. The intelligent reconstruction method for three-dimensional wind speed profile based on deep learning according to claim 5, characterized in that, The total loss of the student model is: ； In the formula, The supervised loss is the true value. Distillation loss of intermediate features in the teacher model. , The adjustable weighting coefficients represent the supervision loss relative to the true value. Distillation loss of intermediate features with the teacher model The weighting of the two parts of the loss in the total loss is calculated as follows: ； ； In the formula, , These represent the feature maps of the teacher model and the student model at the same data index, respectively. For the true wind profile in the tag set, The horizontal wind speed U component at each height in the vertical direction predicted by the student model. The horizontal wind speed V component at each height in the vertical direction is predicted by the student model.

7. A deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles, characterized in that, The system implements the deep learning-based intelligent reconstruction method for three-dimensional wind speed profiles as described in any one of claims 1-6, and the system comprises: The 10m wind speed and 3D wind field data preparation module is used to prepare wind speed components at a height of 10m in a certain sea area using the WRF numerical model. and the three-dimensional wind speed field in the vertical direction Numerical simulations were conducted to obtain regional wind field data for a certain period of time in the sea area, which was then used as the dataset for subsequent models. As input, As output labels; regional wind field data includes the wind component at 10m above ground level. , ; Three-dimensional wind field components , ; The teacher model building and training module is used to slice the 10m wind speed and three-dimensional wind speed data of the sea area into two-dimensional slices in the spatial dimension, which are used as input data for the teacher model, to build and train the teacher model, and to output intermediate feature maps in the middle of the teacher model. Design a knowledge distillation process module to align the intermediate feature maps output by the teacher model and the student model on their respective networks using intermediate feature distillation, and calculate the error between the intermediate feature maps output by the teacher network and the student network. The student model building and training module is used to slice the 10m wind speed and 3D wind speed data of this sea area into two-dimensional spatial single-point data in a spatial dimension. This data is used as the input data for the student model to build and train the student model. The module uses supervised loss based on error and true values, and distillation loss based on intermediate features with the teacher model. Total loss will result; The model application module based on single-point input is used in the practical application stage after both the teacher model and the student model have been trained, by inputting a specific location and time. The complete three-dimensional wind profile at that location is inferred using the student model.

8. The intelligent reconstruction system for three-dimensional wind speed profiles based on deep learning according to claim 7, characterized in that, The system is mounted on a computer device, which includes at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, it implements the functions of the aforementioned deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles.

9. The intelligent reconstruction system for three-dimensional wind speed profiles based on deep learning according to claim 7, characterized in that, The system is mounted on a computer-readable storage medium that stores a computer program. When the computer program is executed by a processor, it can realize the functions of the aforementioned deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles.

10. The intelligent reconstruction system for three-dimensional wind speed profiles based on deep learning according to claim 7, characterized in that, The system is mounted on an information data processing terminal, which provides a user input interface to implement the functions of the deep learning-based intelligent reconstruction system for three-dimensional wind speed profiles as described above when executed on an electronic device.

Images