Ensembles machine learning for short duration rainfall intensity forecasting
2731-0809
Abstract
Abstract This study introduces an ensemble machine learning (EML) framework to enhance short-duration rainfall-intensity prediction under limited-data conditions. By integrating multiple base learners within a stacking architecture, the proposed approach extends current machine learning applications in hydrological modeling. It provides a flexible, data-driven framework for better approximating nonlinear relationships between maximum daily rainfall and statistically derived rainfall-intensity values beyond conventional statistical assumptions. This study aims to predict rainfall intensity at a 10-year return period for short durations of 60 min (DCH60) and 120 min (DCH120) using an EML framework. The dataset comprises maximum daily rainfall observations from 103 stations distributed across Jawa, Sumatera, and Kalimantan from 2015 to 2024, serving as the independent variables. The proposed EML framework integrates six base learners. It employs k-Nearest Neighbors (kNN) as the meta-learner to improve predictive stability and model nonlinear predictor–target relationships in the prepared dataset. Results show that EML predicts DCH60 with R2 l = 0.983, R2 u = 0.919, RMSEL = 2.168 mm/hr, and RMSEU = 2.187 mm/hr. For DCH120, the model achieves R2 l = 0.994, R2 u = 0.923, RMSEL = 0.856 mm/hr, and RMSEU = 1.341 mm/hr. These findings demonstrate that EML provides an adaptive, data-driven alternative for predicting short-duration rainfall intensity once reference rainfall-intensity values are available from conventional hydrological estimation. The proposed approach supports improved hydrological planning, drainage design, and flood risk mitigation in tropical regions with limited high-resolution rainfall observations.