A regional wind farm group multi-scale meteorological data assimilation and spatio-temporal reconstruction method

Abstract

This invention discloses a method for multi-scale meteorological data assimilation and spatiotemporal reconstruction of regional wind farm clusters, aiming to address the problems of insufficient accuracy in existing wind farm meteorological forecasts and difficulties in fusing multi-source data. This method achieves effective assimilation of multi-source heterogeneous meteorological data by constructing a deep convolutional variational autoencoder, combines heterogeneous graph networks and neural network differential equation techniques to achieve high-resolution spatiotemporal reconstruction of three-dimensional wind fields, and introduces physical constraints to ensure the physical consistency of the output results. Experimental results show that the method of this invention can significantly improve the accuracy of wind farm meteorological forecasts, providing reliable meteorological data support for the refined operation and management of wind farm clusters.

Description

Technical Field

[0001] This invention belongs to the field of meteorological data processing and wind farm operation and management technology, and specifically relates to a method for multi-scale meteorological data assimilation and spatiotemporal reconstruction of regional wind farm clusters based on deep learning and physical constraints. Background Technology

[0002] As wind farm clusters continue to expand, traditional meteorological forecasting methods have revealed limitations in dealing with complex terrain and multi-scale meteorological features. These limitations include difficulty in fusing multi-source heterogeneous data, lack of multi-scale feature coupling mechanisms, insufficient accuracy in reconstructing spatiotemporal continuity, inadequate physical consistency guarantees, and low computational efficiency. While existing methods utilize machine learning and deep learning to some extent, they still suffer from issues such as single-source data or lack of physical constraints. Therefore, a novel technical solution is urgently needed that can efficiently fuse multi-source heterogeneous meteorological data, accurately reconstruct high-resolution spatiotemporal continuous meteorological fields, and take into account the physical consistency of atmospheric dynamics to meet the forecasting needs of real-time operation of wind farm clusters. Summary of the Invention

[0003] To address the problems existing in the prior art, the present invention aims to provide a method for multi-scale meteorological data assimilation and spatiotemporal reconstruction of regional wind farm clusters. This method can effectively integrate multi-source heterogeneous meteorological data, accurately reconstruct high-resolution three-dimensional wind fields, and ensure the physical consistency of the output results, thereby providing high-precision meteorological data support for regional wind farm clusters.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] A method for multi-scale meteorological data assimilation and spatiotemporal reconstruction of regional wind farm clusters includes the following steps:

[0006] Step 1: Acquire gridded data from numerical weather prediction, data from micro-meteorological observation stations, satellite remote sensing data, wind turbine sensor data, and terrain information;

[0007] Step 2: Construct a deep convolutional variational autoencoder, which includes a multi-scale feature extraction module, a terrain information fusion module, and a physical constraint decoding module;

[0008] The multi-scale feature extraction module is configured to use convolutional neural networks with different convolutional kernels to extract large-scale weather background field features from numerical weather forecast background fields and small-scale local meteorological disturbance features from micro-meteorological observation station data, respectively. The output features of the two types of networks are weighted and merged through a feature fusion layer with an attention mechanism, and the weights of large-scale and small-scale features are adaptively adjusted to achieve effective combination of information at different scales, providing rich multi-level feature representations for subsequent decoding and spatiotemporal reconstruction.

[0009] The terrain information fusion module is used to fuse terrain information with meteorological latent variables output by the encoder. It assigns different weights to the fusion features according to the complexity of the terrain, increasing the weight in complex areas and decreasing the weight in flat areas, thereby highlighting the regulatory role of terrain in the key influence of the meteorological field.

[0010] The physical constraint decoding module is configured to decode the fused meteorological latent variables into a three-dimensional meteorological analysis field, and to embed atmospheric dynamic physical constraints during the decoding process.

[0011] The deep convolutional variational autoencoder is used to assimilate the data from the micro-meteorological observation station with the background field of numerical weather prediction to generate a three-dimensional meteorological analysis field.

[0012] Step 3: Construct a heterogeneous graph network using wind turbine locations, micro-meteorological observation station locations, and terrain grid points as nodes. Calculate the inverse distance weight based on the spatial distance between nodes and the wind field correlation coefficient weight based on historical observation data. Sum the two weighted values ​​together to determine the connection weights between nodes. Use the physical quantity values ​​of the three-dimensional meteorological analysis field obtained in Step 2 at each node location as graph node features. Establish a graph neural network model combining graph attention mechanism and neural ordinary differential equations. Use the adaptive step-size solver of the neural ordinary differential equations to simulate the dynamic process of node state evolution over time with the wind field, achieving high-resolution spatiotemporal reconstruction of the three-dimensional wind field.

[0013] Step 4: Using the high-resolution 3D wind field generated in Step 3 as input to the graph neural network model, the actual observation data at the corresponding time is used as the training target, and the laws of conservation of mass and momentum are introduced as physical constraints. These are added as loss function terms to the training process of the graph neural network model in Step 3 to further train and fine-tune the graph neural network model, thereby outputting a 3D meteorological field with better physical consistency and higher accuracy.

[0014] Preferably, the physical constraint decoding module in step 2 adopts a gated loop unit structure and uses the terrain information as a conditional input to control the feature fusion ratio at different scale levels during the decoding process. This structure can adaptively adjust the feature fusion strategy under different terrain conditions, enhancing the influence of small-scale features in complex terrain areas and emphasizing the contribution of large-scale features in flat areas, thereby improving the accuracy and adaptability of the model to wind field reconstruction under different terrain conditions.

[0015] Preferably, the deep convolutional variational autoencoder in step 2 is trained using the following composite loss function:

[0016] L total =L recon +λ1L KL +λ2L phys

[0017] Among them, L recon To minimize the difference between the 3D meteorological analysis field and the micrometeorological observation station data, a reconstruction loss is defined; L KL KL divergence loss is used to normalize the distribution of meteorological latent variables; L phys λ1 and λ2 are physical constraint losses used to ensure that the output field conforms to the basic laws of atmospheric dynamics; λ1 and λ2 are equilibrium coefficients.

[0018] Preferably, the graph neural network model in step 4 introduces a spectral loss term L during training. spectral The spectral loss term is calculated by determining the difference in energy spectrum distribution between the reconstructed wind field and the actual wind field in the frequency domain. By introducing the spectral loss term, the energy distribution characteristics of the wind field in the frequency domain can be effectively maintained, thereby enhancing the physical realism of the reconstructed wind field and the accuracy of the turbulent structure.

[0019] Preferably, the physical conservation constraints in step 4 include mass conservation terms and momentum conservation terms. The mass conservation term is achieved by calculating the L2 norm of the wind field divergence; the momentum conservation term is achieved by calculating the residual terms of the Navier-Stokes equations. Introducing physical conservation constraints ensures that the output wind field satisfies the fundamental laws of atmospheric dynamics, significantly improving the physical consistency and reliability of the prediction results.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] This invention combines deep convolutional variational autoencoders, heterogeneous graph networks, and physical constraint mechanisms to achieve efficient fusion of multi-source heterogeneous meteorological data and accurate extraction of cross-scale features. It significantly improves the spatiotemporal reconstruction accuracy under complex terrain conditions and ensures the physical consistency of prediction results by embedding atmospheric dynamic constraints. At the same time, compared with traditional numerical weather prediction methods, it improves computational efficiency and can meet the high-precision and high-efficiency prediction requirements for real-time operation of wind farm clusters. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0024] This embodiment uses a large wind farm cluster in North China as the application scenario. The wind farm cluster includes 8 sub-wind farms with a total installed capacity of 800MW. The terrain is complex, including plains, hills and mountains, with an altitude difference of up to 600 meters, providing an ideal test environment for verifying the effectiveness of the method of the present invention.

[0025] Step 1: Multi-source data acquisition and preprocessing

[0026] The acquisition of multi-source data is the foundation of the entire technical solution. The data types, sources, and characteristics involved in this embodiment are as follows: The data acquisition and preprocessing module covers multi-source observation and simulation data to support the spatiotemporal reconstruction of wind farm clusters. First, gridded data from the National Meteorological Administration's WRF numerical weather prediction system is acquired, using Lambert projection, with a spatial resolution of 9km×9km, a temporal resolution of 3 hours, and a forecast lead time of 72 hours. This data includes elements such as horizontal wind speed (u, v), vertical wind speed (w), temperature, relative humidity, air pressure, precipitation, and cloud cover. Second, real-time observations are acquired from 50 micro-meteorological observation stations within the wind farm area, with a station spacing of 5–15km. Each station is equipped with wind speed and direction sensors at heights of 10m, 70m, and 120m, as well as temperature, humidity, and air pressure sensors, with a data acquisition interval of 10 minutes. Third, MODIS satellite remote sensing data (Terra and Aqua) is acquired, with a spatial resolution of 1km×1km and a daily resolution, mainly including surface temperature, cloud cover, and aerosol data. Fourth, SCADA operation data from 200 wind turbines were collected, including nacelle wind speed and direction, ambient temperature, power output, rotor speed, and generator temperature. Data was collected every minute, and outlier detection and missing value imputation were performed to form a complete time series. Finally, high-precision digital terrain model (DEM) data (30m×30m, accuracy better than 5m) was acquired. Terrain slope, aspect, undulation, roughness, and other parameters were calculated, and a terrain influence matrix was constructed to quantify the disturbance effect of terrain on the wind field, achieving high-precision coupling between wind field dynamics and the terrain environment.

[0027] Step 2: Construction of a depthwise convolutional variational autoencoder

[0028] The deep convolutional variational autoencoder is the core component of this invention, responsible for the effective fusion of multi-source heterogeneous data and the reconstruction of three-dimensional meteorological fields. The encoder includes a multi-scale feature extraction module, a terrain information fusion module, and a physical constraint decoding module.

[0029] (1) Design of multi-scale feature extraction module

[0030] The multi-scale feature extraction module employs a parallel convolutional neural network architecture to process large-scale and small-scale meteorological features separately. The large-scale network targets the numerical weather prediction background field. The first layer uses a 7×7 convolutional kernel with a stride of 2, outputting 64 feature maps to capture large-scale weather systems such as high and low pressure systems and fronts. The second layer uses a 5×5 convolutional kernel with a stride of 1, outputting 128 feature maps to extract medium-scale features. Each layer is followed by batch normalization and ReLU activation, and overfitting is prevented using Dropout (dropout rate 0.3). The small-scale network processes local data from micro-meteorological observation stations. It uses 3×3 convolutions to capture local disturbance details and 1×1 convolutions for feature transformation and dimensionality reduction to maintain sensitivity to local features. The outputs of the two networks are weighted and merged through a feature fusion layer with an attention mechanism, adaptively adjusting the weights of large-scale and small-scale features to achieve effective combination of information at different scales, providing rich, multi-level feature representations for subsequent decoding and spatiotemporal reconstruction.

[0031] (2) Terrain Information Fusion Module

[0032] The terrain information fusion module is an innovative module of this system, aiming to deeply integrate terrain information with meteorological latent variables. Terrain information parameters (height, slope, aspect, etc.) are first mapped into a 64-dimensional terrain feature vector through an embedding layer and a fully connected network to fully express the influence pattern of terrain on the meteorological field. This terrain feature vector is then concatenated with the 256-dimensional meteorological latent variables output by the encoder in terms of feature dimensions to form a 320-dimensional fusion vector. This maintains the relative independence of terrain and meteorological information while facilitating joint processing by the decoder. To further enhance the representation of terrain information, a terrain influence masking mechanism is introduced: different weights are assigned to the fusion features according to the complexity of the terrain, increasing the weight in complex areas such as mountains and decreasing the weight in flat areas, thereby highlighting the regulatory role of terrain in the key influence of the meteorological field.

[0033] (3) Physical constraint decoding module

[0034] The physical constraint decoding module is responsible for restoring the fused high-dimensional features into a three-dimensional meteorological analysis field, and explicitly embedding atmospheric dynamic physical constraints during the decoding process. This module primarily employs a gated recurrent unit (GRU) containing 256 hidden units, possessing excellent temporal modeling capabilities and capable of outputting meteorological field states at multiple time points. A physical constraint layer is superimposed on the GRU output. This layer consists of a fully connected network with 128 neurons, activated using tanh, and its core function is to ensure that the decoding results conform to physical laws such as mass and momentum conservation. Constraints are implemented through a loss function: mass conservation is measured by the squared error of wind field divergence, making the three-dimensional wind field divergence approach zero; momentum conservation is constrained by the balance relationship between the partial derivative of wind speed components with respect to time and the advection term, Coriolis force, etc. Furthermore, terrain information is included as a conditional input in the decoding process, and a gating mechanism adjusts the fusion ratio of large-scale and small-scale features: emphasizing small-scale features in complex terrain areas and highlighting large-scale features in flat areas, thereby improving the model's adaptability to wind field reconstruction under different terrain conditions.

[0035] (4) Design of composite loss function

[0036] The training of the deep convolutional variational autoencoder employs a composite loss function, comprising three main components. The reconstruction loss term uses mean squared error to calculate the difference between the decoded 3D meteorological field and the actual observation data. This term ensures the network can accurately reconstruct the basic characteristics of the meteorological field. Different weights are assigned to errors at different altitudes within the loss function, with higher weights for near-surface layers and relatively lower weights for upper-level layers. This focuses on the near-surface wind field, which is crucial for wind farm operation. The KL divergence loss term normalizes the distribution of latent meteorological variables, ensuring they follow a standard normal distribution. This term follows the basic principles of variational autoencoders, minimizing the KL divergence between the encoded distribution and the prior distribution to achieve regularization constraints on the latent variable space. The physical constraint loss term includes two sub-terms: mass conservation constraint and momentum conservation constraint. The mass conservation constraint is achieved by calculating the L2 norm of the wind field divergence, requiring the reconstructed wind field to satisfy the continuity equation. The momentum conservation constraint is achieved by calculating the balance between the rate of change of wind speed components over time and various external force terms, ensuring the reconstructed wind field conforms to the momentum equation. The three loss terms are combined using a weighted summation, with balance coefficients λ1 and λ2 set to 0.1 and 0.2, respectively. These coefficients were selected based on extensive experimental tuning to achieve an optimal balance between reconstruction accuracy and physical consistency.

[0037] Step 3: Spatiotemporal Reconstruction of Heterogeneous Graph Networks

[0038] The spatiotemporal reconstruction of heterogeneous graph networks is the key technical step in realizing high-resolution three-dimensional wind field reconstruction in this invention.

[0039] (1) Heterogeneous graph construction

[0040] This heterogeneous graph network uses the spatial locations of wind farm clusters as nodes, comprising 1250 nodes in total. These nodes consist of 200 wind turbine location nodes, 50 micro-meteorological observation station location nodes, and 1000 regularly distributed terrain grid nodes. The wind turbine nodes contain both static information such as nacelle height, hub height, geographic coordinates, and terrain elevation, as well as dynamic features such as wind speed, wind direction, and power output. The micro-meteorological observation station nodes record attributes such as location, equipment type, and observation altitude, and also provide real-time data on wind speed, wind direction, temperature, and humidity at multiple altitude levels. The terrain grid nodes are distributed at 1-kilometer intervals, representing terrain parameters such as elevation, slope, aspect, and surface roughness. The connection weights between nodes are calculated using a dual weighting strategy: the spatial distance weight is based on the normalized reciprocal of the Euclidean distance, reflecting the physical law that correlation decreases with increasing distance; the wind field correlation coefficient weight is based on the Pearson correlation coefficient of historical wind speed time series, reflecting the synchronicity of wind field changes at different locations. The final weights are weighted by a spatial distance weight of 0.6 and a correlation coefficient weight of 0.4, taking into account both spatial proximity and wind field physical correlation, thus effectively characterizing the complex spatial-physical coupling features of wind farm clusters.

[0041] (2) Design of graph attention mechanism

[0042] Graph attention mechanisms are used to adaptively calculate the dynamic association strength between nodes and are a core component of heterogeneous graph networks. This heterogeneous graph network core employs a graph attention mechanism to adaptively model the dynamic associations between nodes. It is designed as a multi-head structure containing eight attention heads, each capturing node association patterns from a different perspective, thereby enhancing the model's expressive power. Attention scores are calculated through the inner product of node features and normalized using Softmax to ensure the weight sum is 1. A scaling factor based on the square root of the feature dimension is introduced to prevent excessively large scores from causing gradient vanishing and to improve training stability. To maintain graph locality and reduce computational cost, only node pairs with edges in the original graph participate in attention calculation (adjacency matrix mask constraint). Finally, the node representation is obtained by a weighted average of neighbor features under the attention weights, achieving adaptive fusion of neighborhood information and generating more expressive feature representations.

[0043] (3) Modeling of frequent differential equations

[0044] The ordinary differential equation module is used to simulate the dynamic evolution of wind fields over time, and is a key technology for achieving spatiotemporal reconstruction. The neural network structure adopts a three-layer fully connected design, with 256 neurons in each layer. The first two layers use nonlinear activation functions to characterize complex dynamics, while the third layer uses linear activation to directly output the derivative of the state with respect to time, thus approximating the basic form of the ordinary differential equations through the neural network. Time integration is performed by the Dormand–Prince adaptive step-size solver, which is based on the fifth-order Runge–Kutta framework and has an error control mechanism. It automatically reduces the step size when wind field changes drastically and increases the step size when changes are gradual, balancing stability and efficiency. During training, a random time sampling strategy is further introduced, randomly selecting wind field sequences from different time intervals in each batch for training, thereby enhancing the model's generalization ability to diverse temporal patterns and avoiding overfitting.

[0045] (4) Training strategies and optimization

[0046] The training of the graph neural network adopts a phased strategy: First, the graph attention network is pre-trained using the generated 3D meteorological analysis field as input and measured observations as labels to capture the spatial dependencies between nodes. This phase uses the Adam optimizer (initial learning rate 0.001) and trains for 200 epochs to complete basic feature learning. Subsequently, a neural network frequent differential equation module is introduced for end-to-end joint training, and a combined loss function is designed, including mean squared error loss (ensuring consistency between prediction and observation), physical constraint loss (introducing mass and momentum conservation through soft constraints), and spectral loss (preserving the wind field energy spectrum distribution characteristics, especially satisfying the Kolmogorov-5 / 3 law in the inertial subregion). In the joint training, the learning rate is halved every 50 epochs, for a total of 500 epochs. L2 regularization (weight decay coefficient 1e-4) is used to alleviate overfitting, thereby improving prediction accuracy and generalization ability while ensuring physical consistency.

[0047] Step 4: Fine-tuning of physical constraints and output

[0048] After training in step 3, physical conservation constraints are further introduced to fine-tune the model, improving the physical consistency of the output wind field. The fine-tuning phase uses the high-resolution 3D wind field generated in step 3 as input, with the actual observation data at the corresponding time point as the target. Mass and momentum conservation terms are added to the loss function. The mass conservation term is calculated using the L2 norm of the wind field divergence, and the momentum conservation term is calculated using the residual terms of the Navier-Stokes equations. A small learning rate (1e-5) is used during the fine-tuning process, with 100 training epochs to avoid disrupting the learned feature representations. The final output 3D meteorological field not only closely matches the observation data but also satisfies the fundamental laws of atmospheric dynamics, exhibiting good physical consistency and spatiotemporal continuity.

[0049] Experimental verification and result analysis

[0050] To verify the effectiveness of the method of this invention, a three-month experimental verification was conducted at a wind farm cluster in North China. The experimental results show that:

[0051] (1) In terms of wind speed prediction, the root mean square error (RMSE) of the method of the present invention is reduced by about 28% compared with the traditional WRF method and by 15% compared with the deep learning model without physical constraints.

[0052] (2) In terms of wind direction forecasting, the mean absolute error (MAE) is reduced by about 20%;

[0053] (3) Regarding physical consistency, the mass conservation error is reduced to 10⁻⁵ kg / (m²). 3 The momentum conservation error is on the order of ·s, and is significantly better than that of the comparison method.

[0054] (4) In terms of computing efficiency, it only takes about 5 minutes to complete a 72-hour forecast on a single NVIDIA V100 GPU, which meets the real-time operation requirements of wind farms.

Claims

1. A method for multi-scale meteorological data assimilation and spatiotemporal reconstruction of regional wind farm clusters, characterized in that, Includes the following steps: Step 1: Acquire gridded data from numerical weather prediction, data from micro-meteorological observation stations, satellite remote sensing data, wind turbine sensor data, and terrain information; Step 2: Construct a deep convolutional variational autoencoder, which includes a multi-scale feature extraction module, a terrain information fusion module, and a physical constraint decoding module; The multi-scale feature extraction module is configured to use convolutional neural networks with different convolutional kernels to extract large-scale weather background field features from numerical weather forecast background fields and small-scale local meteorological disturbance features from micro-meteorological observation station data, respectively. The output features of the two types of networks are weighted and merged through a feature fusion layer with an attention mechanism, and the weights of large-scale and small-scale features are adaptively adjusted to achieve effective combination of information at different scales, providing rich multi-level feature representations for subsequent decoding and spatiotemporal reconstruction. The terrain information fusion module is used to fuse terrain information with meteorological latent variables output by the encoder. It assigns different weights to the fusion features according to the complexity of the terrain, increasing the weight in complex areas and decreasing the weight in flat areas, thereby highlighting the regulatory role of terrain in the key influence of the meteorological field. The physical constraint decoding module is configured to decode the fused meteorological latent variables into a three-dimensional meteorological analysis field, and to embed atmospheric dynamic physical constraints during the decoding process. The deep convolutional variational autoencoder is used to assimilate the data from the micro-meteorological observation station with the background field of numerical weather prediction to generate a three-dimensional meteorological analysis field. Step 3: Construct a heterogeneous graph network using wind turbine locations, micro-meteorological observation station locations, and terrain grid points as nodes. Calculate the inverse distance weight based on the spatial distance between nodes and the wind field correlation coefficient weight based on historical observation data. Sum the two weighted values ​​together to determine the connection weight between nodes. The physical quantity values ​​of the three-dimensional meteorological analysis field obtained in step 2 at each node position are used as graph node features. A graph neural network model combining graph attention mechanism and neural ordinary differential equation is established. The adaptive step size solver of the neural ordinary differential equation is used to simulate the dynamic process of node state evolution with wind field time, so as to realize the spatiotemporal reconstruction of high-resolution three-dimensional wind field. Step 4: Using the high-resolution 3D wind field generated in Step 3 as input to the graph neural network model, the actual observation data at the corresponding time is used as the training target, and the laws of conservation of mass and momentum are introduced as physical constraints. These are added as loss function terms to the training process of the graph neural network model in Step 3 to further train and fine-tune the graph neural network model, thereby outputting a 3D meteorological field with better physical consistency and higher accuracy.

2. The method according to claim 1, characterized in that, The physical constraint decoding module adopts a gated loop unit structure and uses the terrain information as a condition input to control the feature fusion ratio at different scale levels during the decoding process.

3. The method according to claim 1, characterized in that, The deep convolutional variational autoencoder described in step 2 is trained using the following composite loss function: L total =L recon +λ1L KL +λ2L phys Among them, L recon To minimize the difference between the 3D meteorological analysis field and the micrometeorological observation station data, a reconstruction loss is defined; L KL KL divergence loss is used to normalize the distribution of meteorological latent variables; L phys λ1 and λ2 are physical constraint losses used to ensure that the output field conforms to the basic laws of atmospheric dynamics; λ1 and λ2 are equilibrium coefficients.

4. The method according to claim 1, characterized in that, The graph neural network model described in step 4 introduces a spectral loss term L during training. spectral The spectral loss term is achieved by calculating the difference between the energy spectrum distribution of the reconstructed wind field and the real wind field in the frequency domain.

5. The method according to claim 1, characterized in that, The physical conservation constraints mentioned in step 4 include mass conservation terms and momentum conservation terms. The mass conservation term is achieved by calculating the L2 norm of the wind field divergence; the momentum conservation term is achieved by calculating the residual terms of the Navier-Stokes equations.

Images