计算机技术与发展2025，Vol.35Issue(5)：129-135,7.DOI:10.20165/j.cnki.ISSN1673-629X.2024.0405
一种用于风场预报订正的双分支时空神经网络模型
A Dual-branch Spatiotemporal Neural Network Model for Wind Field Forecast Correction
摘要
Abstract
The accuracy of weather forecasting is crucial for human activities,especially in the field of wind field forecasting,where precision directly affects the efficiency and safety of industries such as transportation and energy production.Although numerical weather prediction(NWP)has made significant progress,the physical models it relies on still face numerous challenges when simulating wind fields,such as the inability to fully capture the complexity of the Earth's system and the chaotic errors involved.Therefore,data correction plays a pivotal role in improving the accuracy of wind field forecasting.To address this issue,we propose an innovative spatiotemporal recurrent neural network model—DualPhySTRNN.This model transforms the wind field data correction task into a spatiotemporal sequence prediction problem,employing a dual-branch architecture for modeling.One branch imposes physical constraints by introducing partial differential equations,ensuring that the model's output aligns with atmospheric physical laws.The other branch incorporates a Contextual Multi-Scale Enhanced Long Short-Term Memory(CME-LSTM)unit,which fully extracts spatiotemporal information and overcomes the shortcomings of traditional methods that neglect temporal and spatial dependencies.The CME-LSTM module introduces a Contextual Enhancement Block(CEB)and a Multi-Scale Spatiotemporal Expression Block(MSEB)to effectively resolve the issue of inter-frame context information loss and enhance the model's ability to perceive dynamic regions.Experiments conducted on wind field data from Anshan,Liaoning,China,and Central Europe demonstrate that the DualPhySTRNN model significantly outperforms existing mainstream spatiotemporal sequence prediction models in wind field correction tasks.关键词
时空序列预测/风场预报订正/循环神经网络/物理信息神经网络/上下文增强Key words
spatiotemporal sequence prediction/wind field forecast correction/recurrent neural networks/physics-informed neural networks/context enhancement分类
信息技术与安全科学引用本文复制引用
谷占鑫,吴杰,金巍,韩国敬..一种用于风场预报订正的双分支时空神经网络模型[J].计算机技术与发展,2025,35(5):129-135,7.基金项目
中国科学院国家空间科学中心空间天气学国家重点实验室开放性课题(E211AQ1S) （E211AQ1S）
