Hydrological Ensemble Modeling of Juwata Station Maximum Rainfall using FFNN and RNN

Abstract

This study aims to develop a more accurate and reliable hydrological forecasting model for predicting maximum rainfall at Juwata Station in Tarakan, North Kalimantan. The objective is to address the limitations of traditional statistical approaches when handling highly dynamic and non-linear hydrometeorological patterns. To achieve this, we apply an ensemble modeling approach that integrates four neural network architectures: FFNN Backpropagation, FFNN Rprop, Elman RNN, and Jordan RNN. The methods involve training each model independently, evaluating their predictive performance, and combining their outputs using a weighted averaging scheme that assigns greater influence to models with lower error metrics. The analysis focuses on the ability of each architecture to capture non-linear relationships and temporal dependencies within the rainfall data. The findings show that the ensemble model produces consistently lower SMAPE, MSE, and RMSE values compared to any single model or two-model combination, demonstrating superior predictive capability. The novelty of this study lies in its integration of multiple feedforward and recurrent neural network structures into a unified ensemble specifically tailored for tropical rainfall forecasting, where atmospheric variability is high. This ensemble framework offers a more robust foundation for predicting extreme rainfall events, supporting flood risk mitigation and regional infrastructure planning.