Physics-Informed Deep-Learning Prediction of Tropical Cyclone Wind Size in the Western North Pacific

Accurate forecasting of tropical cyclone (TC) size remains a significant challenge in operational meteorology. This study develops a physics-informed long short-term memory (LSTM) model to improve TC size prediction in the Northwest Pacific, leveraging multi-source data, including best track records, global reanalysis datasets, and outputs from an operational regional TC numerical model of the China Meteorological Administration (CMA-TYM). Key predictors, such as TC internal characteristics (maximum wind speed and initial vortex size) and environmental variables (ocean heat content, relative humidity, and vertical wind shear), are diagnostically selected to train a bidirectional LSTM with an attention mechanism for forecasting the radius of maximum wind (RMW) and wind profile parameters. The model achieves a 24-h mean absolute error of 8.3 km for RMW (91% of samples with error < 20 km) and 34.8 km for R17 (78% of samples with error < 50 km). For 120-h forecasts, the R17 relative error stabilizes at 0.1–0.2, exhibiting robustness against sample scarcity and slower error growth than the conventional methods. Compared with CMA-TYM, the LSTM model reduces the 6–48 h forecast errors by 25.8% for a damaging wind radius of 26 m s⁻¹ and 38.5% for a hurricane-force radius of 33 m s⁻¹ while also capturing asymmetric wind-field structures and improving the long-axis wind distribution accuracy during TC intensification. Sensitivity analysis identified the initial vortex size, ocean thermal conditions, and environmental humidity as the dominant drivers of size variability. The proposed framework demonstrates superior stability and resolution in TC size forecasting, providing a valuable tool for disaster risk assessment and early warning systems.