Author(s): Yasaman Afkhami; Isa Ebtehaj; Silvio Jose Gumiere; Francois Anctil; Hossein Bonakdari

Linked Author(s): François Anctil, Hossein Bonakdari

Keywords: GR4J; Improved Extreme Learning Machine (IELM); Machine learning; Quebec; Rainfall-runoff model; Streamflow modelling

Abstract: Accurate and timely daily streamflow simulation is essential for informed decision-making in various sectors, such as water resources management, flood prediction, drought mitigation, and hydroelectric power generation. This study compares the performance of the Improved Extreme Learning Machine (IELM) as a machine learning-based model and GR4J (i. e., Genie Rural a 4 parametres Journalier in French) as a conceptual hydrological-based model. The IELM is a rapid, non-tuned, single-layer feed-forward neural network that is integrated with an iterative process to address the issue of random parameter assignment in the conventional form of the extreme learning machine. The GR4J was combined with a snow module to accurately represent snow accumulation and melt processes in hydrological modeling, as well as with a genetic algorithm to optimize the parameters of the GR4J and snow module. Data from the Saint-Charles hydrometric station in Quebec, Canada, were taken on a daily basis. Approximately 8000 samples were gathered from January 01,2001, to October 31,2022. The statistical indices suggest that both models exhibit satisfactory generalizability, as the disparities between the training and testing indices are negligible across all statistical measures. Moreover, the IELM (R = 0.94, NSE = 0.87, PBIAS (%) = 0.76, NRMSE = 0.46, MAE = 0.21), outperformed the GR4J model (R = 0.88, NSE = 0.73, PBIAS (%) = 16.28, NRMSE = 0.68, MAE = 2.82) across all applied statistical measures. Comparing these two models in terms of seasons reveals that IELM demonstrates superior performance in all seasons except winter.

DOI: https://doi.org/10.3850/978-90-833476-1-5_iahr40wc-p0236-cd

Year: 2023