Three-dimensional wind field reconstruction method and system based on deep learning multi-source meteorological data

Abstract

The invention discloses a three-dimensional wind field reconstruction method and system based on deep learning multi-source meteorological data. The method comprises the following steps: collecting wind field data through a Doppler laser radar and a wind profile radar; based on an MLP as a trunk architecture, introducing position coding, and constructing a neural field model; and performing zero-sample super-resolution prediction on the wind field data based on the neural field model to generate a reconstructed three-dimensional wind field. According to the method, the complex space-time correlation in the wind field can be adaptively learned, and the space-time continuity of the reconstruction result is ensured through overlapped boundary processing. The three-dimensional wind field reconstruction method and system based on the deep learning multi-source meteorological data can be widely applied to the technical field of three-dimensional wind field reconstruction.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional wind field reconstruction technology, and in particular to a three-dimensional wind field reconstruction method and system based on deep learning multi-source meteorological data. Background Technology

[0002] The acquisition and fusion of wind field observation data is a crucial foundation for meteorological science, urban environmental planning, and renewable energy development. Traditional observation methods, such as weather stations, lidar, and ultrasonic anemometers, typically record wind field data at fixed heights and time intervals. However, in actual monitoring networks, instrument malfunctions, operational errors, and environmental interference often lead to missing or abnormal data. Furthermore, significant differences exist among various observation devices in terms of spatiotemporal resolution, accuracy, and coverage, making the fusion of multi-source heterogeneous wind data an extremely challenging task. In recent years, artificial intelligence technologies, particularly deep learning methods, have offered new solutions to this field due to their powerful nonlinear modeling capabilities. These methods can learn complex spatiotemporal relationships from multi-source heterogeneous data, reconstructing high-resolution, continuous wind field information. Their technical approach shares some similarities with the super-resolution problem in computer vision. However, wind field data fusion faces two unique challenges: firstly, unlike image super-resolution, wind field observations typically lack high-resolution, high-precision "ground truth" data as supervisory signals; secondly, random equipment malfunctions or data loss result in a large number of missing values ​​in the model input, making traditional super-resolution methods difficult to apply directly. Currently, typical methods for wind field reconstruction include machine learning interpolation algorithms such as Support Vector Machine (SVM), Radial Basis Function (RBF), Random Forest (RF), Kriging (GP), and Gradient Boosting Decision Tree (XGBoost). However, while these methods demonstrate some effectiveness in specific scenarios, they have inherent limitations when dealing with complex urban wind environments. Methods like SVM and RBF, based on kernel function mapping mechanisms, have nonlinear expressive power limited by the predefined kernel function form, making it difficult to adaptively capture multi-scale turbulence features in wind fields. Ensemble learning methods such as Random Forest and Gradient Boosting Decision Tree, while capable of handling nonlinear relationships, have weak generalization ability for continuous spatiotemporal coordinates and often produce over-smoothing results when modeling high-frequency turbulence components. Gaussian process regression, while providing a probabilistic prediction framework, has computational complexity that increases cubically with the amount of data, making it unsuitable for real-time modeling of large-scale wind field data. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide a three-dimensional wind field reconstruction method and system based on deep learning multi-source meteorological data. This method can adaptively learn the complex spatiotemporal relationships in the wind field and ensure the spatiotemporal continuity of the reconstruction results through overlapping boundary processing.

[0004] The first technical solution adopted in this invention is a three-dimensional wind field reconstruction method based on deep learning multi-source meteorological data, comprising the following steps: Wind field data were collected using Doppler lidar and wind profiler radar. Based on MLP as the backbone architecture, positional encoding is introduced to construct a neural field model; Zero-sample super-resolution prediction of wind field data is performed based on a neural field model to generate a reconstructed three-dimensional wind field.

[0005] Furthermore, the step of acquiring wind field data using Doppler lidar and wind profiler radar specifically includes: Three-dimensional wind speed and wind direction angle data were acquired using Doppler lidar. Horizontal wind field data are collected using wind profiler radar; Wind field data is constructed by combining three-dimensional wind speed data, wind direction angle data, and horizontal wind field data.

[0006] Furthermore, the step of constructing the neural field model specifically includes: A continuous mapping relationship from spatiotemporal coordinates to wind field parameters is constructed. MLP is selected as the backbone architecture, position encoding is introduced, and Fourier feature mapping is applied to the preprocessed input to construct a neural field model.

[0007] Furthermore, the step of performing zero-sample super-resolution prediction of wind field data based on a neural field model to generate a reconstructed three-dimensional wind field specifically includes: Considering the periodicity of wind direction, a trigonometric transformation is performed on the wind field data to decompose the wind direction into components on two independent axes, resulting in wind field data in unit vector form. A multi-task learning strategy is adopted, a shared layer is set to capture the low-level feature representations of multiple tasks, and a task-specific layer is set to output the corresponding target. The wind field data in unit vector form is predicted to obtain the predicted wind direction. The predicted wind direction is represented by a combination of two prediction components, and is corrected according to the difference in wind direction to obtain the corrected predicted wind direction; The corrected predicted wind direction is divided into multiple independent segments, and the mean squared error is used as the loss function to train the neural field model, thus obtaining the trained neural field model. Zero-sample super-resolution prediction of wind field data is performed based on the trained neural field model to generate a reconstructed three-dimensional wind field.

[0008] Furthermore, the expression for the predicted wind direction is as follows: ; In the above formula, Indicates the predicted wind direction. and They represent the predicted and Quantity.

[0009] Furthermore, the specific expression for the corrected predicted wind direction is as follows: ; In the above formula, Indicates the observed wind direction.

[0010] The second technical solution adopted in this invention is: a three-dimensional wind field reconstruction system based on deep learning multi-source meteorological data, comprising: The first module is used to collect wind field data through Doppler lidar and wind profiler radar; The second module is used to construct a neural field model based on MLP as the backbone architecture and by introducing position encoding. The third module is used to perform zero-sample super-resolution prediction of wind field data based on the neural field model, and generate the reconstructed three-dimensional wind field.

[0011] The beneficial effects of the method and system of this invention are as follows: This invention acquires wind field data through Doppler lidar and wind profiler radar; further, based on MLP as the backbone architecture, it introduces position encoding to construct a neural field model, and uses the neural field framework to construct a continuous mapping function from spatiotemporal coordinates to wind data. Through MLP, end-to-end nonlinear modeling is achieved, which can adaptively learn the complex spatiotemporal correlations in the wind field. Furthermore, Fourier feature mapping is introduced as a position encoding technique to map the original spatiotemporal coordinates to a high-dimensional frequency domain space, which effectively enhances the ability of MLP to capture multi-scale turbulence features and solves the spectral bias problem of traditional MLP. Finally, based on the neural field model, zero-sample super-resolution prediction is performed on the wind field data to generate the reconstructed three-dimensional wind field. The spatiotemporal continuity of the reconstruction results is ensured through overlapping boundary processing. Attached Figure Description

[0012] Figure 1 This is a flowchart of the steps of the three-dimensional wind field reconstruction method based on deep learning multi-source meteorological data of the present invention; Figure 2 This is a structural block diagram of the three-dimensional wind field reconstruction system based on deep learning multi-source meteorological data of the present invention; Figure 3 This is a schematic diagram of the neural field-based multi-source wind data fusion method NF-MW provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of a Fourier feature mapping example provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of wind speed and direction modeling provided in a specific embodiment of the present invention. Detailed Implementation

[0013] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0014] First, it's important to note that the wind field reconstruction method based on neural fields exhibits significant advantages. Neural fields establish a continuous mapping function from spatiotemporal coordinates to wind field parameters using deep neural networks. This end-to-end modeling approach can adaptively learn complex nonlinear relationships within the wind field. In particular, by introducing positional encoding techniques such as Fourier eigenmaps, neural fields effectively mitigate the spectral bias problem of traditional multilayer perceptrons (MLPs), enabling them to capture both large-scale wind field structures and reconstruct subtle turbulent fluctuations. This optimization of spectral characteristics allows the model to accurately reproduce high-frequency details in the wind field while maintaining global consistency.

[0015] Based on this, this invention proposes a deep learning-based multi-source wind data fusion technique called NF-MW (Neural Fields for Multi-source Winds). NF-MW utilizes neural networks to map spatiotemporal coordinates to wind data output, enabling simultaneous training of wind data collected by various instruments. The model combines two key modules—Fourier function mapping and piecewise implementation—to enhance its nonlinearity and accurately characterize atmospheric boundary laminar flow. Furthermore, wind direction is decomposed into orthogonal components to ensure its periodicity. This method demonstrates extremely high accuracy in wind data reconstruction, potentially significantly enhancing urban wind environment analysis in densely populated cities and supporting data-driven decision-making for sustainable urban planning.

[0016] Reference Figure 1 This invention provides a method for reconstructing three-dimensional wind fields based on deep learning multi-source meteorological data. The method includes the following steps: S100: Collects wind field data using Doppler lidar and wind profiler radar; Specifically, three-dimensional wind speed and wind direction angle data are collected using Doppler lidar; horizontal wind field data are collected using wind profiler radar; and wind field data are constructed by combining the three-dimensional wind speed data, wind direction angle data, and horizontal wind field data.

[0017] From March 28, 2025 to April 28, 2025, wind field data for a certain city were collected using Doppler lidar and wind profiler radar. The Doppler lidar had an observation height range of 52.0–493.6 meters, a vertical resolution of 25.9 meters, and a temporal resolution of 1 minute, enabling precise measurement of three-dimensional wind speed and wind direction. The wind profiler radar covered an observation height range of 150–510 meters, with a vertical resolution of 120 meters and a temporal sampling interval of 5 minutes, primarily acquiring horizontal wind field data. We selected wind data collected by the Doppler lidar at heights of 207.8 m and 389.7 m as the validation dataset, and combined the remaining observation data from the Doppler lidar with all observation data from the wind profiler radar to form the training dataset.

[0018] S200, based on MLP as the backbone architecture, introduces position encoding to construct a neural field model; Specifically, a continuous mapping relationship from spatiotemporal coordinates to wind field parameters is constructed, MLP is selected as the backbone architecture, position encoding is introduced, and Fourier feature mapping is applied to the preprocessed input to construct a neural field model.

[0019] In this embodiment, NF-MW establishes a spatiotemporal coordinate system. Wind field parameters The continuous mapping relationship allows for the prediction of wind data at any altitude. Specifically, NF-MW at altitude... and time Combined input and corresponding altitude time output A mapping function was established between them: ; in Indicates the observed wind speed (m / s) or wind direction (°).

[0020] We chose an MLP as the backbone to perform the coordinate-based wind data modeling described above. The MLP model contains multiple fully connected (FC) layers to learn complex feature representations. Each FC layer It can be represented as: ; in yes -th Output of the FC layer; yes -th FC layer input (also -th FC layer output). and They are -Th FC layer's learnable weights and biases; It is a non-linear activation function. In this method, we will... Specify the ReLU activation function to promote non-negative output.

[0021] Simultaneously, this method introduces positional encoding to enhance the nonlinearity of the MLP. To effectively capture high-frequency wind behavior, Fourier feature mapping (a form of positional encoding) has proven to be a solution for controlling the frequency learning capability of the MLP. In this study, we apply Fourier feature mapping to the preprocessed input. The core idea of ​​this method is to apply a trigonometric transformation to map the input to a high-frequency space. We represent the input (height and time) as... Positional encoding maps it to an encoded feature vector. This vector belongs to a higher-dimensional space. As shown below: ; in This represents the hyperparameter with the highest frequency used to adjust the smoothness of the encoded features in the mapping. The positional encoding module enhances the dimension of the input. The original input vector has a dimension of... The feature dimension after position encoding is .For example, Figure 4 This demonstrates the Fourier feature map (sine transform), where the coordinates The three encoded features are projected onto a high-dimensional space.

[0022] S300 performs zero-sample super-resolution prediction of wind field data based on a neural field model to generate a reconstructed three-dimensional wind field.

[0023] Specifically, considering the periodicity of wind direction, a trigonometric transformation is performed on the wind field data to decompose the wind direction into components on two independent axes, resulting in wind field data in unit vector form. A multi-task learning strategy is adopted, setting a shared layer to capture the low-level feature representations of multiple tasks, while setting task-specific layers to output corresponding targets, predicting the wind direction from the unit vector form of the wind field data. The predicted wind direction is represented by a combination of the two predicted components and corrected according to the wind direction difference, resulting in a corrected predicted wind direction. The corrected predicted wind direction is segmented into multiple independent segments, and the mean squared error is used as the loss function to train the neural field model, resulting in a trained neural field model. Based on the trained neural field model, zero-sample super-resolution prediction is performed on the wind field data to generate a reconstructed 3D wind field.

[0024] In this embodiment, as Figure 5 As shown, in this method, the output is... This indicates wind speed or direction. Wind speed modeling is straightforward, while wind direction modeling requires considering its periodicity. For example, the interpolation between 350° and 10° is actually 20°. We have added a detailed discussion and solutions, as follows: Due to the periodicity of wind direction, it is necessary to further convert the wind direction into a unit vector.

[0025] ; in Indicates wind direction; and This indicates a trigonometric transformation. Unit vector output ensures the continuity of the interpolation while avoiding output normalization.

[0026] Wind direction interpolation is handled by transforming it into two unit vectors, each corresponding to a component of the wind direction on an independent axis. Modeling each component independently ignores the inherent correlation between them and exponentially increases the computational cost of training, although it offers implementation simplification. However, they are semantically related in their predictions because their values ​​are subject to strict mathematical constraints (i.e., Due to limitations, they will change interdependently. For example, a [missing information - likely a specific feature or characteristic] used for... The predictor can use this information to guide another predictor. The predictor ensures that the predicted components are semantically consistent with the aforementioned mathematical constraints. Multi-task learning aims to share some low-level information, thereby improving accuracy and avoiding additional computational costs. Multi-task learning sets up a shared layer to capture the low-level feature representations of multiple tasks, and then sets up multiple task-specific layers to output the corresponding targets.

[0027] The predicted wind direction can be represented as follows: ; in Indicates the predicted wind direction (°). and They represent the predicted and Quantity. It depends on the wind direction. Make corrections.

[0028] ; in This indicates the observed wind direction. The above function can limit wind direction variations. Between 0° and 180°.

[0029] We further introduced a piecewise implementation to ensure the model's nonlinearity. A municipal meteorological bureau provided wind data from March 28, 2015 to April 28, 2025 to evaluate the performance of the proposed data fusion model. Although the location-encoded MLP can handle all spatiotemporal wind data, the high temporal dimension significantly impacts the model's learning efficiency.

[0030] To accurately capture wind field characteristics, a segmented approach has proven to be a good solution in practice. Specifically, we divide the wind field data sequence into multiple independent segments, each representing one day's data (1440 minute coordinate points along the time direction). Each segment corresponds to an independent MLP, used to train and reconstruct the wind field in the subdomain. To ensure the continuity of the reconstruction results along the segments, we extend both ends of each segment by one hour. Because this segmented implementation supports parallel computing and incurs lower training cycles, the computational cost does not increase significantly.

[0031] Next, we will train an MLP for each segment. The overall modeling model is as follows: Figure 5 As shown. Input includes time. and height They will be passed separately Fourier feature maps are used to encode each input feature to expand it, and then these are concatenated into a single encoded feature. The encoded feature is then processed by an MLP to output the final target, i.e., wind speed or wind direction. Each MLP contains 10 standard fully connected layers with ReLU activation functions, each with 128 channels. All these models are implemented in PyTorch and trained on an iGame GeForce RTX 3090 24G GPU. Optimization is performed using the Adam algorithm with a learning rate of 0.001. Training is conducted with a batch size of 256 and runs for 200 epochs. The loss function is mean squared error (MSE).

[0032] After training, we can perform zero-shot super-resolution to predict wind data at arbitrary heights and times. We iteratively select the model's input from the test dataset and then output wind data to evaluate the Fourier feature map and MLP. The superiority of this model represents its potential for generating high-fidelity reference databases. When deployed on a single GPU (GeForce RTX 3090 24GGPU), the proposed method completes a single-segment training (i.e., one day's worth of wind data) in less than 1 minute (approximately 43 seconds).

[0033] In summary, this invention proposes a multi-source wind data fusion method based on neural fields (NF-MW). This method achieves high-precision three-dimensional wind field reconstruction by constructing a continuous mapping function from spatiotemporal coordinates to wind data. Figure 3 As shown, this method comprises four core functional modules: a data preprocessing module, a neural field modeling module, a wind field reconstruction module, and a validation and evaluation module. The data preprocessing module is responsible for the acquisition and standardization of multi-source heterogeneous wind field data; the neural field modeling module constructs the mapping relationship between spatiotemporal coordinates and wind field parameters; the wind field reconstruction module realizes continuous modeling of high-resolution three-dimensional wind fields; and the validation and evaluation module performs visualization analysis and accuracy verification of the reconstruction results.

[0034] The embodiments of the present invention have the following distinguishing features compared with the prior art: 1) A continuous mapping function from spatiotemporal coordinates to wind data is constructed using the Neural Fields framework. End-to-end nonlinear modeling is achieved through MLP. This architecture can adaptively learn the complex spatiotemporal correlations in the wind field.

[0035] 2) Introducing Fourier feature mapping as a position coding technique maps the original spatiotemporal coordinates to a high-dimensional frequency domain space, which effectively enhances the ability of MLP to capture multi-scale turbulence features and solves the spectral bias problem of traditional MLP.

[0036] 3) In view of the periodic characteristics of wind direction, an innovative orthogonal decomposition method is adopted to convert the wind direction into sine and cosine components for independent modeling. Finally, the wind direction is reconstructed through the arctangent function, thus avoiding the problem of wind direction discontinuity.

[0037] 4) A segmented implementation strategy is proposed, which divides the entire spatiotemporal domain into multiple subdomains and assigns independent neural networks for modeling. This significantly reduces computational complexity and supports parallel training. At the same time, the spatiotemporal continuity of the reconstruction results is ensured through overlapping boundary processing.

[0038] Therefore, the embodiments of the present invention have the following advantages compared with the prior art: 1) By combining the neural field framework with Fourier feature mapping, accurate modeling of multi-scale features of wind fields is achieved. Compared with traditional machine learning methods (such as SVM, RBF, etc.), it has stronger nonlinear expression capabilities and can capture large-scale and small-scale turbulence features at the same time.

[0039] 2) The orthogonal decomposition and segmented implementation strategy, which addresses the issues of wind direction periodicity and time dimension limitations, effectively solves the discontinuity problem in wind direction modeling of existing methods and the efficiency bottleneck in large-scale wind field data processing, providing a high-precision and high-efficiency solution for urban wind environment analysis.

[0040] Reference Figure 2 A 3D wind field reconstruction system based on deep learning multi-source meteorological data includes: The first module 201 is used to collect wind field data through Doppler lidar and wind profiler radar; The second module 202 is used to construct a neural field model based on MLP as the backbone architecture and by introducing position encoding. The third module 203 is used to perform zero-sample super-resolution prediction of wind field data based on the neural field model, and generate the reconstructed three-dimensional wind field.

[0041] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0042] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A three-dimensional wind field reconstruction method based on deep learning multi-source meteorological data, characterized in that, Includes the following steps: Wind field data were collected using Doppler lidar and wind profiler radar. Based on MLP as the backbone architecture, positional encoding is introduced to construct a neural field model; Zero-sample super-resolution prediction of wind field data is performed based on a neural field model to generate a reconstructed three-dimensional wind field.

2. The three-dimensional wind field reconstruction method based on deep learning multi-source meteorological data according to claim 1, characterized in that, The step of acquiring wind field data using Doppler lidar and wind profiler radar specifically includes: Three-dimensional wind speed and wind direction angle data were acquired using Doppler lidar. Horizontal wind field data are collected using wind profiler radar; Wind field data is constructed by combining three-dimensional wind speed data, wind direction angle data, and horizontal wind field data.

3. The three-dimensional wind field reconstruction method based on deep learning multi-source meteorological data according to claim 2, characterized in that, The step of constructing the neural field model specifically includes: A continuous mapping relationship from spatiotemporal coordinates to wind field parameters is constructed. MLP is selected as the backbone architecture, position encoding is introduced, and Fourier feature mapping is applied to the preprocessed input to construct a neural field model.

4. The three-dimensional wind field reconstruction method based on deep learning multi-source meteorological data according to claim 3, characterized in that, The step of performing zero-sample super-resolution prediction of wind field data based on a neural field model to generate a reconstructed three-dimensional wind field specifically includes: Considering the periodicity of wind direction, a trigonometric transformation is performed on the wind field data to decompose the wind direction into components on two independent axes, resulting in wind field data in unit vector form. A multi-task learning strategy is adopted, a shared layer is set to capture the low-level feature representations of multiple tasks, and a task-specific layer is set to output the corresponding target. The wind field data in unit vector form is predicted to obtain the predicted wind direction. The predicted wind direction is represented by a combination of two prediction components, and is corrected according to the difference in wind direction to obtain the corrected predicted wind direction; The corrected predicted wind direction is divided into multiple independent segments, and the mean squared error is used as the loss function to train the neural field model, thus obtaining the trained neural field model. Zero-sample super-resolution prediction of wind field data is performed based on the trained neural field model to generate a reconstructed three-dimensional wind field.

5. The three-dimensional wind field reconstruction method based on deep learning multi-source meteorological data according to claim 4, characterized in that, The expression for the predicted wind direction is as follows: ； In the above formula, Indicates the predicted wind direction. and They represent the predicted and Quantity.

6. The three-dimensional wind field reconstruction method based on deep learning multi-source meteorological data according to claim 5, characterized in that, The specific expression for the corrected predicted wind direction is as follows: ； In the above formula, Indicates the observed wind direction.

7. A three-dimensional wind field reconstruction system based on deep learning multi-source meteorological data, characterized in that, Includes the following modules: The first module is used to collect wind field data through Doppler lidar and wind profiler radar; The second module is used to construct a neural field model based on MLP as the backbone architecture and by introducing position encoding. The third module is used to perform zero-sample super-resolution prediction of wind field data based on the neural field model, and generate the reconstructed three-dimensional wind field.

Images