Improving short-term rainfall prediction with PSO-FD-LSTM deep learning

Short-term rainfall forecasting using long short-term memory (LSTM) networks is often constrained by sub-optimal hyperparameter selection and data non-stationarity. To address these limitations, we propose the PSO-FD-LSTM model, which integrates first-order differencing (FD) to enhance data stationarity and employs particle swarm optimization (PSO) for adaptive LSTM hyperparameter tuning. Evaluated using data from Fushanluqu station, China, the framework consistently outperformed standard SVR, LSTM, FD-LSTM, and PSO-LSTM models across various lead times. Specifically, for 6 h forecasts, the proposed model achieved an R2 of 0.9563 (a 15.8% improvement over the standard LSTM), while significantly reducing the MAE and RMSE to 0.3217 and 1.1591, respectively. These results demonstrate that combining physical data transformation with algorithmic optimization effectively overcomes traditional modelling bottlenecks, providing a highly accurate and robust tool for real-time short-term rainfall forecasting.