Typhoon track and temporal forecasting of wind and rain based on a multifactor coupling model and LSTM networks

Abstract

This study proposes an integrated typhoon forecasting framework that sequentially combines a multi-factor physical model with a Long Short-Term Memory (LSTM) network. The physical component handles track projection via typhoon genesis indices, multi-layer steering flows, and β-drift corrections, and characterizes post-landfall wind decay through the Kaplan attenuation model; the LSTM network independently models temporal precipitation using features derived from physical outputs; and Dynamic Time Warping (DTW) quantifies track prediction fidelity. Validated on Typhoons Bebinca (2024) and Lekima (2019) in the Northwest Pacific, the model achieves an average track error of 45.2 km (DTW distance 0.46), with maximum relative errors of 8.6% for post-landfall central wind speed and 9.8% for average precipitation. The framework demonstrates complementary benefits of combining physical interpretability with data-driven nonlinear fitting, offering a technical reference for refined Northwest Pacific typhoon forecasting and disaster mitigation.