Neural network river discharge forecasters: an empirical investigation of hidden unit processing functions based on two different catchments

This paper extends previous attempts at detecting physical process representation inside a neural network hydrological forecasting solution. The hidden unit processing functions inside two nonlinear autoregressive neural network river flow forecasting models are examined using two river catchments: the River Brosna in the Republic of Ireland and the Blue Nile in East Africa. Multiple linear regression was found to be a useful tool to investigate the internal relationships that exist between the hidden unit outputs and the observed discharge record. The residual plots, from the regression analysis, were used to explore the role and function of the different processing units and one hidden unit was discovered to have captured most of the input-output relationship. This unit provided a good linear approximation to the observed discharges. The other hidden units exhibited much weaker relationships with respect to observed discharge. Their main influence on model outputs was instead found to be a complex and integrated nonlinear correction factor that had differential impacts on linear unit approximations at low-to-intermediate and upper magnitude flood events.