River Flow Prediction Using an Integrated Approach

Abstract

River flow predictions are needed in many water resource management activities. Hydrologists have relied on individual techniques such as time series, conceptual, or artificial neural networks (ANNs) to model the complex rainfall-runoff process in the past. These techniques, when used individually, provide reasonable accuracy in modeling and forecasting river flow. This paper presents an integrated approach for river flow prediction in an attempt to achieve better forecast accuracy. Specifically, three different models are presented for daily river flow prediction: a time series model of autoregressive type, a nonlinear conceptual model, and an integrated model. The conceptual model uses the Green-Ampt method to model infiltration, time area method to translate rainfall input in time, and a nonlinear reservoir for flood routing. The integrated model uses conceptual, ANN, genetic algorithm, data-decomposition, and model-fusion techniques. The data derived from the Kentucky River basin were employed to calibrate and validate all the models. A wide variety of standard performance indices was used to evaluate model performance. The performance of the time series model was found to be the worst and the conceptual model was found to perform reasonably well. The integrated model performed the best, demonstrating a need for developing innovative hybrid models capable of exploiting advantages of individual techniques in order to achieve improved accuracies in short-term river flow forecasting.