Neural network approach to output updating for the physically-based model of the Upper Thames River watershed

Abstract

This study presents an output updating procedure for the deterministic physically-based model of the Upper Thames River watershed, Ontario, Canada. In addition to streamflow and rainfall, this procedure uses as inputs meteorological variables not employed in the model calibration. The main hydrological processes involved in transformation of rainfall into runoff are mathematically expressed using a set of key variables. Therefore, some of the available meteorological variables may be of limited value during the calibration that predominantly relies on a large range of flow hydrographs for obtaining the optimum state variables and parameters of the model. In this study, the Bayesian regularisation neural network technique is coupled with the physically-based model to provide more accurate flood flow simulation for a wide range of flood flow event hydrographs pertinent to the hydrometeorological environment. The artificial neural network is capable of generating good generalisation results after complex input-output mapping.