Enhancing daily runoff prediction: A hybrid model combining GR6J-CemaNeige with wavelet-based gradient boosting technique

Abstract

Hydrological modeling is essential for understanding and managing water resources, predicting flood events, and assessing the impacts of climate change on hydrological cycles. Previous research has shown the potential of machine learning (ML) models in hydrological modeling, but there remains a gap in effectively integrating these models with specific hydrological processes. This study addresses the challenges of runoff simulation in cold regions by systematically integrating Gradient Boosting Model (GBM) models with a hydrological process-based model (namely Génie Rural à 6 paramètres Journalier (GR6J) model coupled with CemaNeige snow module (GR6J-CemaNeige)) to improve hydrological modeling approaches. Four various schemes were examined for combining GBM with GR6J-CemaNeige, including production store combinations, unit hydrograph combinations, routing store concepts, and snowmelt and snowpack combinations. The GR6J-CemaNeige model achieved a Kling-Gupta Efficiency (KGE) of 0.775 and a Nash-Sutcliffe Efficiency (NSE) of 0.686 in the test sections, establishing a process-based baseline model for runoff simulation. The production store combinations yielded KGE values ranging from 0.722 to 0.745 and NSE from 0.601 to 0.614, while unit hydrograph combinations achieved KGE values of 0.8 and 0.804 and NSE values 0.702 and 0.705 during the test sections. The routing store combinations presented promising results with KGE values ranging from 0.805 to 0.822 and NSE values ranging from 0.71 to 0.734 for the test sections. Notably, the snowmelt and snowpack combinations achieved KGEs ranging from 0.743 to 0.759 and NSEs ranging from 0.641 to 0.666 during the test sections. The application of signal processing techniques, specifically Maximal Overlap Discrete Wavelet Transform (MWT) and Multiresolution Analysis (MRA), further improved runoff simulation accuracy across various hydrological components. The best MWT results were derived from the unit hydrograph scenario (MWT-GBM7), achieving a KGE of 0.881 and a NSE of 0.816 in the test section, demonstrating the technique’s effectiveness in capturing complex snow-related processes. For MRA, the routing store scenario (MRA-GBM9) produced the best results with a KGE of 0.881 and a NSE of 0.788 in the test section, highlighting the method’s capability to enhance the representation of runoff timing and distribution. The consistent improvement across different hydrological components suggests that the hybrid approach successfully captures complex interactions within the watershed.