A Dual-Branch LSTM Model for Short-Term Rainfall Forecasting Integrating GNSS-Derived PWV and Surface Meteorological Parameters

Accurate short-term rainfall forecasting is essential for disaster mitigation. Although numerical weather prediction models are widely used, their application to short lead times is constrained by computational demands. Data-driven approaches provide an efficient alternative. To better exploit atmospheric water vapor information, this study develops a dual-branch long short-term memory (LSTM) model that integrates Global Navigation Satellite System (GNSS)-derived precipitable water vapor (PWV) with surface meteorological parameters for rainfall forecasting. The model processes historical rainfall and meteorological variables through separate branches. Historical rainfall characterizes precipitation persistence, while PWV, PWV variation (ΔPWV), PWV rate of change (ΔtPWV), and air temperature describe atmospheric moisture evolution and thermodynamic conditions before rainfall. The model was evaluated using hourly observations from 18 GNSS-collocated meteorological stations in Taiwan collected during 2018–2019 and compared with a rainfall history-based LSTM baseline model. Results show that the proposed model achieved accuracies of 89–91% and recalls of 88–90% for 1–3 h forecasts. Its advantages became more evident for longer lead times, with Recall and Threat Score increasing by 6–11% and 4–8%, respectively, for 2–3 h forecasts. These findings demonstrate that integrating GNSS-derived PWV with surface meteorological parameters can improve short-term rainfall forecasting.