Daily rainfall forecast based on multi-station observation data in Medan City

Abstract

This study develops a multi-horizon daily rainfall forecasting model using the Long Short-Term Memory (LSTM) deep learning method, based on multi-station Automatic Weather Station (AWS) data in Medan City. Ten-minute AWS data from multiple stations (2021–2024) were merged and time-synchronized (UTC), followed by a quality control process including physical range checks, rate-of-change filtering, inter-variable consistency checks, and spike detection. Missing values were addressed using linear interpolation for short gaps and Multiple Imputation by Chained Equations (MICE) for longer gaps. Predictor features were constructed from weather parameters (temperature, humidity, pressure, wind, radiation), aggregated to an hourly scale, and reshaped into input time windows for LSTM. A two-layer LSTM model (128–64 units, 0.3 dropout, Adam optimizer) was trained to predict daily rainfall up to five days ahead. Evaluation metrics, including RMSE, MAE, POD, FAR, and CSI (with rainfall threshold ≥1 mm/day), indicated strong model performance: for instance, RMSE was below 10 mm/day for 1–3 day horizons, with POD above 0.80 and FAR below 0.20. The LSTM model outperformed conventional statistical models, yielding an accuracy improvement of approximately 30–40%. These findings highlight the potential of high-resolution AWS-based automatic forecasting systems to support hydrometeorological disaster mitigation in tropical urban areas.