Cross-urban-area wind field rapid forecasting method based on meta learning and dynamic data assimilation

Abstract

The invention discloses a meta learning and dynamic data assimilation-based cross-urban wind field rapid forecasting method, which realizes efficient modeling of a complex city canopy effect through a cross-urban feature standardization encoder in combination with Li DAR point cloud and a Fourier neural operator (FNO), and remarkably improves the wind field forecasting precision. Through a transferable element learning framework, geometric disturbance and physical disturbance strategies are combined, so that the model can quickly adapt to a dynamically changing meteorological background field, and the response time of an emergent weather event is greatly shortened. Through a dynamic data assimilation engine, a GPU is adopted to accelerate a CFD solver and ensemble Kalman filtering (EnKF), unification of high-precision and efficient calculation is achieved, and the real-time performance and robustness of the model are remarkably improved.

Description

Technical Field

[0001] The present invention deeply intersects the field of meteorological forecasting and artificial intelligence, specifically a method for rapid forecasting of cross-city wind fields based on meta-learning and dynamic data assimilation. Background Art

[0002] 1. Existing technical solutions:

[0003] 1) Wind speed correction method based on single geographic information (GIS) feature

[0004] Patent CN113435344A uses geographic information such as elevation and slope to correct wind speed. This approach may be effective in areas with relatively simple terrain, but in complex urban environments, it neglects the impact of the three-dimensional form of buildings on turbulence, resulting in insufficient prediction accuracy. Specifically, turbulence characteristics are complex in areas with dense high-rise buildings, and a single geographic feature cannot capture the interactions between buildings and the wake effects.

[0005] 2) Neural network models with long temporal dependencies

[0006] Patent CN114492214A uses an autoformer neural network to capture long-term temporal dependencies. While this model performs well in processing time series data, its static parameters are unable to adapt to sudden weather events. For example, during a severe convective weather test in Nanjing, the model's response delay reached 42 minutes, resulting in a 17% failure rate for drone route adjustments. This demonstrates that existing AI models have significant shortcomings in their dynamic response capabilities and are unable to adapt promptly to sudden changes in the meteorological background.

[0007] 3) Fixed coupled WRF-CFD nested model

[0008] Patent CN112950056A simulates wind fields using a fixed coupling of WRF and CFD. While this model improves simulation accuracy to a certain extent, it lacks a closed-loop data assimilation mechanism, resulting in low computational efficiency. Specifically, a single 72-hour forecast requires 36 hours of use across 128 core nodes in a supercomputing center, failing to meet city-level real-time requirements. Furthermore, the fixed coupling approach lacks flexibility and is difficult to adapt to the diverse needs of different urban areas.

[0009] 2. Defects of existing technology:

[0010] 1) Unable to handle complex urban canopy effects

[0011] Single geographic information features cannot capture the turbulent characteristics of areas densely populated with high-rise buildings. The three-dimensional shape of buildings significantly influences the wind field, especially in areas with dense high-rise buildings, where the nonlinear effects of turbulence are more pronounced. Existing technologies ignore these factors, resulting in significant variations in prediction accuracy across different urban areas and failing to meet the needs of cross-city wind forecasting.

[0012] 2) Insufficient dynamic response capability

[0013] Existing AI models often use offline training, leading to exponentially higher prediction errors when the meteorological background undergoes sudden changes. For example, a temperature gradient reversal can cause errors to soar by 35%. This dynamic response blind spot renders existing technologies ineffective in responding to sudden weather events, making them unable to provide reliable real-time forecasts for scenarios such as dynamic route planning for low-altitude economic drones.

[0014] 3) The contradiction between computational efficiency and accuracy

[0015] While high-precision LES (Large Eddy Simulation) can provide detailed turbulence information, its temporal resolution is limited to one hour, which cannot meet real-time requirements. While lightweight LSTM models offer high computational efficiency, their turbulent kinetic energy error exceeds 25%, resulting in insufficient prediction accuracy. This conflict between computational efficiency and accuracy makes it difficult for existing technologies to achieve a good balance in practical applications. Summary of the Invention

[0016] The purpose of the present invention is to solve the problems mentioned in the above background technology by introducing a cross-city feature normalization encoder, a transferable meta-learning framework and a dynamic data assimilation engine.

[0017] To solve the above technical problems, the present invention provides a method for rapid cross-city wind farm forecasting based on meta-learning and dynamic data assimilation, which is characterized by comprising the following steps:

[0018] Step 1: Cross-city feature standardization encoder

[0019] ① 3D architectural topology analysis

[0020] The building complex graph structure is constructed based on LiDAR point cloud. The node attributes include: height H, volume V, facade roughness C d , vegetation coverage rate G r ,

[0021] The edge weight of the graph is defined as the building distance D ij Function: Among them H avg is the average building height to quantify the intensity of the wake effect;

[0022] ②Surface roughness field modeling

[0023] The Fourier neural operator (FNO) is used to process high-resolution 0.1m satellite remote sensing data. The spatial continuous distribution of roughness z0 is learned in the frequency domain. The attention mechanism is introduced to focus on road-building interface areas such as street canyons. The weight formula is:

[0024]

[0025] ③Dynamic weight optimization

[0026] Design a self-supervised task: By randomly masking 20% of the building blocks, the encoder is forced to learn the implicit association between geometry and flow field. The weight coefficients α, β, γ, δ are dynamically generated by the gating network:

[0027] α=σ(W a [GNN(X),FNO(Z)])

[0028] Where σ is the sigmoid function, which realizes the adaptive distribution of feature importance;

[0029] Step 2: Transferable Meta-Learning Framework

[0030] ①Multi-task perturbation strategy

[0031] Geometric perturbation: Randomly scale, translate, and rotate building heights to simulate layout differences in different urban areas;

[0032] Physical perturbation: Introducing the random source term S into the Navier-Stokes equations ω , covering Reynolds number Re=10 4 ~10 6 flow pattern changes;

[0033] ② Second-order optimization acceleration

[0034] The inner loop uses Hessian-Free optimization to avoid explicit calculation of the second-order derivative matrix:

[0035]

[0036] The outer loop uses the Lookahead optimizer, and the parameter update rules are:

[0037] θ fast =θ+β(θ slow -θ)

[0038] Effectively alleviate the oscillation problem in meta-learning;

[0039] ③Knowledge distillation

[0040] The CFD high-fidelity simulation results are used as the teacher model, and the KL divergence is used to constrain the output distribution of the student model:

[0041]

[0042] Step 3: Dynamic Data Assimilation Engine

[0043] ① Uncertainty quantification

[0044] Monte Carlo Dropout is sampled 50 times to calculate the confidence interval of the wind speed forecast:

[0045]

[0046] Dynamic threshold setting: Based on historical error distribution, the POT (Peaks Over Threshold) model is used to adaptively adjust the trigger threshold;

[0047] ②GPU-accelerated CFD solver

[0048] CUDA-based parallel computing for unstructured grids and algebraic multigrid AMG preprocessing are used to increase the speed of solving RANS equations by 12 times.

[0049] Time step adaptive control: Dynamically adjust Δt through CFL conditions to balance stability and efficiency;

[0050] ③ Ensemble Kalman Filter

[0051] Design a localized covariance matrix to avoid spurious correlations between distant observations:

[0052]

[0053] Among them L c =500m is the relevant length scale;

[0054] Parameter-state joint assimilation: Simultaneously update the model weights θ and flow field states u, v, w.

[0055] Furthermore, the geometric perturbation in step 2 imposes a random scaling of ±30% on the building height, a maximum translation of 50m, and a rotation of 0-180°.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. Improved modeling efficiency: The cross-city feature standardization encoder of the present invention significantly simplifies the modeling process of complex urban canopy effects by using LiDAR point cloud and FNO technology, reducing modeling time by more than 60%;

[0058] 2. High computational accuracy: In a cross-continental test set, the mean absolute error (MAE) of wind farm predictions in new urban areas was reduced from 2.31m / s compared to existing technologies to 0.74m / s, with a simulation error of less than 4%;

[0059] 3. Strong dynamic adaptability: The transferable meta-learning framework combines geometric and physical perturbation strategies to enable the model to quickly adapt to sudden weather events, reducing response latency from 42 minutes in existing technologies to 8.2 minutes, and improving dynamic adaptability by 3.5 times.

[0060] 4. Efficient computing and real-time performance: The dynamic data assimilation engine uses GPU-accelerated CFD solvers and EnKF, increasing computing efficiency from 3.4 TFlops of existing technology to 21.6 TFlops, a 63% increase in computing efficiency.

[0061] 5. Excellent generalization ability: Through meta-learning and knowledge distillation technology, the present invention demonstrates excellent generalization performance in cross-continental tests, reducing the cross-continental prediction error from 142% of the existing technology to 9.7%;

[0062] 6. Full Process Automation: This invention integrates data collection, feature encoding, and dynamic assimilation, reducing manual intervention by 85%. In particular, in a drone emergency obstacle avoidance scenario, the system predicted a sudden strong gust of wind between buildings nine minutes in advance, successfully avoiding three potential collisions and significantly improving operational safety. DETAILED DESCRIPTION

[0063] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0064] In Example 1, the present invention provides a technical solution: a method for rapid forecasting of cross-city wind farms based on meta-learning and dynamic data assimilation, comprising the following steps:

[0065] Step 1: Cross-city feature standardization encoder

[0066] ① 3D architectural topology analysis

[0067] The building complex graph structure is constructed based on LiDAR point cloud. The node attributes include: height H, volume V, facade roughness C d , vegetation coverage rate G r ,

[0068] The edge weight of the graph is defined as the building distance D ij Function: Among them H avg is the average building height to quantify the intensity of the wake effect;

[0069] ②Surface roughness field modeling

[0070] The Fourier neural operator (FNO) is used to process high-resolution 0.1m satellite remote sensing data. The spatial continuous distribution of roughness z0 is learned in the frequency domain. The attention mechanism is introduced to focus on the road-building interface (such as street canyons). The weight formula is:

[0071]

[0072] ③Dynamic weight optimization

[0073] Design a self-supervised task: By randomly masking 20% of the building blocks, the encoder is forced to learn the implicit association between geometry and flow field. The weight coefficients α, β, γ, δ are dynamically generated by the gating network:

[0074] α=σ(W α [GNN(X),FNO(Z)])

[0075] Where σ is the sigmoid function, which realizes the adaptive distribution of feature importance;

[0076] Step 2: Transferable Meta-Learning Framework

[0077] ①Multi-task perturbation strategy

[0078] Geometric perturbation: Randomly scale, translate, and rotate building heights to simulate layout differences in different urban areas;

[0079] Physical perturbation: Introducing the random source term S into the Navier-Stokes equations ω , covering Reynolds number Re=10 4 ~10 6 flow pattern changes;

[0080] ② Second-order optimization acceleration

[0081] The inner loop uses Hessian-Free optimization to avoid explicit calculation of the second-order derivative matrix:

[0082]

[0083] The outer loop uses the Lookahead optimizer, and the parameter update rules are:

[0084] θ fast =θ+β(θ slow -θ)

[0085] Effectively alleviate the oscillation problem in meta-learning;

[0086] ③Knowledge distillation

[0087] The CFD high-fidelity simulation results are used as the teacher model, and the KL divergence is used to constrain the output distribution of the student model:

[0088]

[0089] Step 3: Dynamic Data Assimilation Engine

[0090] ① Uncertainty quantification

[0091] Monte Carlo Dropout is sampled 50 times to calculate the confidence interval of the wind speed forecast:

[0092]

[0093] Dynamic threshold setting: Based on historical error distribution, the POT (Peaks Over Threshold) model is used to adaptively adjust the trigger threshold;

[0094] ②GPU-accelerated CFD solver

[0095] CUDA-based parallel computing for unstructured grids and algebraic multigrid AMG preprocessing are used to increase the speed of solving RANS equations by 12 times.

[0096] Time step adaptive control: Dynamically adjust Δt through CFL conditions to balance stability and efficiency;

[0097] ③ Ensemble Kalman Filter

[0098] Design a localized covariance matrix to avoid spurious correlations between distant observations:

[0099]

[0100] Among them L c =500m is the relevant length scale;

[0101] Parameter-state joint assimilation: Simultaneously update the model weights θ and flow field states u, v, w.

[0102] This method is not only applicable to dynamic route planning for low-altitude economical drones and real-time monitoring of urban microclimates, but can also be extended to scenarios such as pollutant dispersion simulation and building wind load assessment. Through multimodal data fusion and an adaptive model architecture, it achieves a paradigm shift in meteorological services from static analysis to dynamic response.

[0103] This invention achieves efficient modeling of complex urban canopy effects and significantly improves wind field prediction accuracy by combining a cross-city feature standardization encoder with LiDAR point clouds and Fourier neural operators (FNO). Through a transferable meta-learning framework, combined with geometric perturbation and physical perturbation strategies, the model can quickly adapt to dynamically changing meteorological background fields, significantly shortening the response time to sudden weather events. Through a dynamic data assimilation engine, a GPU-accelerated CFD solver and an ensemble Kalman filter (EnKF) are used to achieve the unity of high precision and efficient computing, significantly improving the real-time and robustness of the model.

[0104] Example 2, in a cross-city wind field rapid forecasting method based on meta-learning and dynamic data assimilation according to Example 1, the geometric perturbation in step 2 applies a random scaling of ±30% to the building height, a maximum translation of 50m, and a rotation of 0-180°.

[0105] Example 3, experimental verification

[0106] 1. Dataset: We collaborated with Bettair (30 cities in Europe) and UO-Wind (15 cities in North America) to build a cross-continental test set.

[0107] 2. Benchmark model: Comparison with seven mainstream methods including WRF-CMAQ, DeepCFD, and GraphCast.

[0108] index WRF-CMAQ DeepCFD The present invention MAE in new urban area (m / s) 2.31 1.89 0.74 Burst response delay (min) 45 28 8.2 Computing energy efficiency (TFlops) 0.17 3.4 21.6 Generalization (across continents) fail Error +142% Error +9.7%

[0109] The cross-city feature standardization encoder of the present invention: constructs the building complex graph structure through LiDAR point cloud, combines the Fourier neural operator (FNO) to process high-resolution satellite remote sensing data, uniformly describes the attributes such as the height, volume, facade roughness and vegetation coverage of the building complex, quantifies the intensity of the wake effect, and realizes efficient modeling of the complex urban canopy effect. Transferable meta-learning framework: adopts geometric perturbation (random scaling, translation, rotation) and physical perturbation (introduction of random source terms) strategies, combined with second-order optimization acceleration (Hessian-Free optimization and Lookahead optimizer) and knowledge distillation technology, so that the model can quickly adapt to the layout differences of different urban areas and significantly improve the dynamic response capability. Dynamic data assimilation engine: Based on GPU-accelerated CFD solver and ensemble Kalman filter (EnKF), through Monte Carlo Dropout sampling and time step adaptive control, it achieves the unity of high precision and efficient computing. At the same time, through uncertainty quantification and parameter-state joint assimilation, it further improves the real-time and robustness of the model. Real-time error correction and optimization: Root mean square error (RMSE) triggers iterative optimization, dynamically adjusting model weights and flow field states to ensure high accuracy and stability in wind farm forecasts. Through these innovations, this invention achieves a paradigm shift from static analysis to dynamic response, providing an efficient and accurate solution for rapid cross-city wind farm forecasting.

[0110] The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes based on the technical solution and invention concept of the present invention within the scope disclosed by the present invention, which falls within the scope of protection of the present invention.

Claims

1. A method for rapid cross-city wind farm forecasting based on meta-learning and dynamic data assimilation, characterized by: The following steps are involved: Step 1: Cross-city feature standardization encoder ① 3D architectural topology analysis The building complex graph structure is constructed based on LiDAR point cloud. The node attributes include: height H, volume V, facade roughness C d , vegetation coverage rate G r , The edge weight of the graph is defined as the building distance D ij Function: Among them, H avg is the average building height to quantify the intensity of the wake effect; ②Surface roughness field modeling The Fourier neural operator (FNO) is used to process high-resolution 0.1m satellite remote sensing data. The spatial continuous distribution of roughness z is learned in the frequency domain. The attention mechanism is introduced to focus on the road-building interface. The weight formula is: ③Dynamic weight optimization Design a self-supervised task: By randomly masking 20% of the building blocks, the encoder is forced to learn the implicit association between geometry and flow field. The weight coefficients α, β, γ, δ are dynamically generated by the gating network: α=σ(W α [GNN(X),FNO(Z)]) Where σ is the sigmoid function, which realizes the adaptive distribution of feature importance; Step 2: Transferable Meta-Learning Framework ①Multi-task perturbation strategy Geometric perturbation: Randomly scale, translate, and rotate building heights to simulate layout differences in different urban areas; Physical perturbation: introducing the random source term S in the Navier-Stokes equations ω , covering Reynolds number Re=10 4 ~10 6 flow pattern changes; ② Second-order optimization acceleration The inner loop uses Hessian-Free optimization to avoid explicit calculation of the second-order derivative matrix: The outer loop uses the Lookahead optimizer, and the parameter update rules are: i fast =θ+β(θ slow -i) Effectively alleviate the oscillation problem in meta-learning; ③Knowledge distillation The CFD high-fidelity simulation results are used as the teacher model, and the KL divergence is used to constrain the output distribution of the student model: Step 3: Dynamic Data Assimilation Engine ① Uncertainty quantification Monte Carlo Dropout is sampled 50 times to calculate the confidence interval of the wind speed forecast: Dynamic threshold setting: Based on historical error distribution, the POT model is used to adaptively adjust the trigger threshold; ②GPU-accelerated CFD solver CUDA-based parallel computing for unstructured grids and algebraic multigrid AMG preprocessing are used to increase the speed of solving RANS equations by 12 times. Time step adaptive control: Dynamically adjust Δt through CFL conditions to balance stability and efficiency; ③ Ensemble Kalman Filter Design a localized covariance matrix to avoid spurious correlations between distant observations: Among them L c =500m is the relevant length scale; Parameter-state joint assimilation: Simultaneously update the model weights θ and flow field states u, v, w.

2. The method for rapid cross-city wind farm forecasting based on meta-learning and dynamic data assimilation according to claim 1, characterized in that: The geometric perturbation in step 2 applies a random scaling of ±30% to the building height, a maximum translation of 50m, and a rotation of 0-180°.

Images