Scholarlink

<p>Accurate weather forecasting is vital for various societal and economic activities, yet it remains a challenging task due to the complex spatio-temporal dynamics of atmospheric processes. Traditional numerical weather prediction models are computationally intensive and often struggle to capture intricate local and global dependencies inherent in meteorological data. Recent deep learning approaches have made strides in spatio-temporal forecasting but typically focus on either local or global patterns, leading to suboptimal performance. In this paper, we propose a novel framework named CTWPre that integrates Convolutional Long Short-Term Memory (ConvLSTM) networks and Transformer architectures in a dual-branch model to simultaneously capture local and global spatio-temporal dependencies. The outputs of these branches are fused and fed into a Partial Differential Equation (PDE) module that utilizes symbolic regression to automatically discover the underlying physical equations governing the data. This integration enables our model to not only improve prediction accuracy but also provide insights into the control of complex dynamical systems. Experiments on meteorological datasets demonstrate that our approach outperforms state-of-the-art methods in both accuracy and computational efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Physics-aware dual-branch architectures for accurate weather predictions

  • 期刊论文

  • 发表日期：

    2025年2月17日

查看全文
  • Cuilian Zhang,
  • Weijun Zhou

摘要

Accurate weather forecasting is vital for various societal and economic activities, yet it remains a challenging task due to the complex spatio-temporal dynamics of atmospheric processes. Traditional numerical weather prediction models are computationally intensive and often struggle to capture intricate local and global dependencies inherent in meteorological data. Recent deep learning approaches have made strides in spatio-temporal forecasting but typically focus on either local or global patterns, leading to suboptimal performance. In this paper, we propose a novel framework named CTWPre that integrates Convolutional Long Short-Term Memory (ConvLSTM) networks and Transformer architectures in a dual-branch model to simultaneously capture local and global spatio-temporal dependencies. The outputs of these branches are fused and fed into a Partial Differential Equation (PDE) module that utilizes symbolic regression to automatically discover the underlying physical equations governing the data. This integration enables our model to not only improve prediction accuracy but also provide insights into the control of complex dynamical systems. Experiments on meteorological datasets demonstrate that our approach outperforms state-of-the-art methods in both accuracy and computational efficiency.