Comparative analysis of deep learning models for multi-horizon rainfall forecasting in flood-prone tropical highlands

Abstract

Accurate rainfall forecasting is critically important for the effective mitigation of floods and the management of water resources in tropical highland regions that serve as principal upstream catchment areas for major reservoirs. Heavy periods of rainfall in the upper watershed can exceed a reservoir’s capacity for storage and release, thus contributing to the recurrence of flooding in downstream urban areas and their surrounding regions. This study addresses this practical model-selection problem by systematically comparing Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Fusion Transformer (TFT), and AutoTFT with multiple daily forecasting horizons for rainfall prediction. Historical rainfall records were combined with meteorological variables to develop the forecasting models. Model performance was evaluated across short- and medium-term forecasting horizons using Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). The results indicate that AutoTFT consistently achieved the lowest MAE across all forecasting horizons, ranging from 6.75 mm (1-day) to 7.79 mm (14-day). At the 3-day and 9-day horizons, AutoTFT also produced the lowest MSE (201.36 and 215.91 mm²) and RMSE (14.19 and 14.69 mm), demonstrating superior predictive accuracy. At the 7-day and 14-day horizons, TFT slightly outperformed AutoTFT in terms of RMSE (14.66 and 14.88 mm, respectively), although AutoTFT maintained the lowest MAE. Meanwhile, LSTM achieved the lowest RMSE (12.79 mm) at the 1-day horizon, indicating competitive performance for very short-term forecasting. Overall, the transformer-based models, particularly AutoTFT and TFT, consistently outperformed the recurrent architectures at medium- and long-range forecasting horizons, highlighting their potential as reliable and generalizable approaches for rainfall forecasting in flood-prone tropical highlands and supporting more effective early warning systems and water resource management.