Ensemble learning of catchment-wise optimized LSTMs enhances regional rainfall-runoff modelling - case study: Basque Country, Spain

Abstract

Accurate rainfall-runoff modeling is crucial for effective water resources management and planning, especially in flash catchments prone to rapid floods. This study investigates the performance of ensemble learning methods applied to regionally optimized deep learning models, specifically long short-term memory (LSTM) networks, for enhanced hydrological prediction. Three ensemble approaches were developed based on optimized regional hyperparameter settings: catchment-wise, top-10 regional, and K-means clustering selected configurations. These networks were trained, and the median of their simulations on the test set was considered the final prediction for each ensemble. The final predictions were then evaluated against observed data. Our findings show that ensemble learning methods consistently outperform conventional single-configuration approach of selecting the best regional setting in all locations, especially in catchments with prediction complexity or anthropogenic footprints. The catchment-wise ensemble demonstrated the highest prediction accuracy and robustness, highlighting the importance of tailoring network configurations to the unique characteristics of individual catchments. The findings highlight the potential of ensemble learning to significantly improve hydrological forecasts and inform better decision-making in water resources management. Specifically, this research demonstrates how ensemble learning of catchment-wise configurations can overcome limitations in regional hydrological predictions by deep learning models, addressing the "uniqueness of the place" paradigm.