Coordinate Attention Based Wide & Deep Fusion Network for Typhoon Track Prediction

Guangxin He, Jiadian Wang, Jie Feng, Jingjia Luo, Xun Li

Tropical Cyclone Research and Review · 17 April 2026 · Fudan University (FuXi), Nanjing University of Information Science & Technology (NUIST)

Paper

In one sentence. Wide and Deep LSTM and ConvGRU network with coordinate attention predicts typhoon tracks using CMA best-track data and ERA5 fields for 2020–2024.

Abstract

Accurate forecasting of the tracks of tropical cyclones is of crucial importance for effective disaster risk management in coastal regions. These storms are recognized as one of nature's most hazardous weather phenomena. However, traditional typhoon track prediction methods face certain limitations in handling complex meteorological data and multi-dimensional features. To improve prediction accuracy, this paper proposes an improved Wide & Deep framework. This framework utilizes a Stacked Two-layer Long Short-Term Memory (S2-LSTM) encoder in the Wide branch to model the sequential features of typhoon tracks, and incorporates a Multi-ConvGRU in the Deep branch to handle three-dimensional meteorological data. Following a dynamic/static feature decomposition, the S2-LSTM encoder processes a subset of dynamic features, effectively integrating typhoon position, wind speed, and rate-of-change information from multiple time steps to better capture the temporal evolution patterns of typhoon movement. Furthermore, a Coordinate Attention (CA) module is introduced, which captures positional dependencies along the vertical and horizontal dimensions within the feature maps through a factorized average pooling operation. This enhances the model's ability in identifying key areas within atmospheric fields, thus ensuring improved spatial precision in typhoon track forecasting. The model's efficacy was assessed with the 2020–2024 typhoon track dataset published by the China Meteorological Administration (CMA) and the ERA5 reanalysis product maintained by the European Centre for Medium-Range Weather Forecasts (ECMWF). The findings indicate how the two introduced enhancements to the model lead to higher precision for deep learning-based forecasts to a certain extent.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.