Forecasting Concurrent Flows in a River System Using ANNs

Abstract

Applications of artificial neural networks (ANNs) in forecasting flow rates at multiple sections in a river system are presented. Model formulations are based on learning characteristics of the actual and fractional storage variations in river reaches during unsteady flow. Multilayer perceptrons (MLP), MLPs with memory, time delay neural networks (TDNNs), and multiple gamma memory neural networks in three model forms are used to forecast flow rates in Tar River Basin, United States. Model performances are evaluated in terms of statistical criteria, RMS error, and coefficient of efficiency. Maximum RMS error resulted for the models are less than 6.50% of the respective observed mean value. A coefficient of efficiency value of more than 0.95 for the models indicates satisfactory performances. Results presented in this paper depict flow variations corresponding to implicitly specified storage variations and demonstrate applicability of the ANNs in real time flow forecasting for multiple sections in a basin obeying continuity principle.