Abstract

Machine learning (ML) techniques have gained popularity in watershed management and planning due to their accurate forecasting capabilities. The study aims to evaluate and compare the performance of various machine learning models for monthly runoff forecasting in the Thale Sap Songkhla Basin and assess their effectiveness in improving runoff prediction using predicted rainfall as input. The data utilized in this study included hydrological and meteorological data (i.e., runoff, rainfall, relative humidity, air temperature, and wind speed), as well as large-scale climate indicators (LSCI), namely the Southern Oscillation Index, sea surface temperature, and the Indian Ocean Dipole Mode Index. Statistical performance metrics such as the correlation coefficient (r), mean absolute error (MAE), root mean square error (RMSE), and overall index (OI) between the observed and forecasted data were employed to evaluate and compare the performances of the ML models. The results indicate that meteorological and LSCI, particularly sea surface temperature, are crucial for monthly rainfall and runoff forecasting. The comparative results showed that the SVR-rbf model exhibited better performance in monthly rainfall forecasting at 5 out of 12 stations. The validation results also revealed that the SVR-rbf and RF models exhibited the highest performance for forecasting runoff at 4 stations each out of 8, for the Thale Sap Songkhla basin at most stations. Adjusting additional parameters for RF models, such as bagSizePercent, improved model performance. Additionally, using forecasted rainfall as an input for runoff prediction improved model performance at 7 out of 8 stations, with accuracy gains of up to 42.34%.