Monthly River Forecasting Using Instance-Based Learning Methods and Climatic Parameters

Surface water resources management relies on the river flow in the region which, in turn, depends on numerous factors, resulting in the complexity of predicting the runoff. In this study, data-driven methods have been used to identify the relation between the river flow and regional climatic parameters and the teleconnection indexes. To achieve this, three nonlinear models of artificial neural networks, namely, generalized feedforward neural networks (GFNNs), Jordan-Elman network (JEN), and k-nearest neighbor (KNN), have been used to model monthly flow in a period of 30 years. The sensitivity analysis of input data was done using gamma test, and upon determination of the effective input parameters, modeling was done in four scenarios. The results reveal that among data-driven models, JordanElman neural networks, compared with the other two models, show higher capabilities. On average, the JEN model, in comparison with the KNN and GFNN models, shows 23.4 and 23.04% less errors, respectively. Applying climatic parameters with remote sources, for instance, North Atlantic Oscillation and East Pacific/North Pacific, can enhance the efficiency of GFNN and JEN models. (C) 2017 American Society of Civil Engineers.