Evaluation of Data-Driven Methods for Hydrological Modeling: A Case Study of the Etobicoke Creek Watershed

Abstract

In the past two decades, data-driven modeling has become a popular approach for different modeling tasks. This paper presents an evaluation of the performance of five widely used data-driven approaches (i.e., generalized linear model, lasso regression, support vector machine, neural networks, and random forest) for the modeling of the Etobicoke Creek watershed in Ontario, Canada. The models are built with eleven years of meteorological and hydrometric data from local stations, and the performance is examined by the Nash-Sutcliffe efficiency coefficient, coefficient of determination, mean absolute percentage error, and root mean squared error. The results show all the models are able to generate acceptable predictions and random forest has the highest accuracy. This study can provide support for the selection of hydrological modeling approaches in future studies.