Hybrid streamflow modelling using machine learning and multi-model combination

Abstract

Global hydrological models (GHMs) enable global estimation of freshwater availability, but their uncertainties and limitations hinder precise predictions. Multi-model combination (MMC) is a promising solution that combines the outputs of numerous hydrological models to create an ensembled output that surpasses the individual hydrological models. Moreover, the use of Machine Learning (ML) as a hybrid post-processing strategy is growing in popularity. However, there is a need to combine these two methods and investigate their performance in streamflow predictions. In this study, we demonstrate that using Random Forest (RF) as a nonlinear MMC approach significantly enhances streamflow forecasts when multiple global hydrological models' outputs are combined. In streamflow forecasting, the RF-MMC method outperforms individual models and linear MMC approaches, demonstrating its potential. In addition, incorporating catchment attributes improved the generalizability of the RF-MMC method when tested on a river basin that was not in the training set. Significant potential exists for the application of RF-MMC to generate accurate streamflow forecasts, thereby providing valuable support for water resource management, flood mitigation, and decision-making processes. Future research can investigate additional machine learning algorithms and incorporate additional variables to improve the predictive ability and generalizability of MMC strategies in hydrological modelling.