Potential of Support-Vector Regression for Forecasting Stream Flow

Abstract

Stream flow is an important input for hydrology studies because it determines the water variability and magnitude of a river. Water resources engineering always deals with historical data and tries to estimate the forecasting records in order to give a better prediction for any water resources applications, such as designing the water potential of hydroelectric dams, estimating low flow, and maintaining the water supply. This paper presents three soft-computing approaches for dealing with these issues, i.e. artificial neural networks (ANNs), adaptive-neuro-fuzzy inference systems (ANFISs), and support vector machines (SVMs). Telom River, located in the Cameron Highlands district of Pahang, Malaysia, was used in making the estimation. The Telom River's daily mean discharge records, such as rainfall and river-level data, were used for the period of March 1984-January 2013 for training, testing, and validating the selected models. The SVM approach provided better results than ANFIS and ANNs in estimating the daily mean fluctuation of the stream's flow.