Comparative study of machine learning methods and GR2M model for monthly runoff prediction

Abstract

Monthly runoff time-series estimation is imperative information for water resources planning and development projects. This article aims to comparatively investigate the applicability of machine learning (ML) methods (i.e., Random Forest (RF), M5 model tree (M5), Support Vector Regression with polynomial kernel function (SVR-poly), and Support Vector Regression with the radial kernel function (SVR-rbf)) and the GR2M model for simulating the monthly runoff hydrograph. The models experimented at six runoff stations in Thailand’s Southern basin. Four performance criteria, including Nash-Sutcliffe Efficiency (NSE), Correlation Coefficient (r), Overall Index (OI), and Combined Index (CI), were utilized for model performance comparison. The finding results revealed that in stations with a low correlation coefficient (r) between input and output data sets, ML algorithms showed superior performance to GR2M. In particular, SVR-rbf showed outstanding performance over other methods. It expressed that SVR-rbf could manage the problem of low-quality data and simulate monthly runoff under limited available data.

Keywords

• GR2M
• M5 Model tree
• Random forest
• Runoff
• Support vector regression