Input–output dynamic neural networks simulating inflow–outflow phenomena in an urban hydrological basin

Abstract

In this paper neural networks have been studied as a tool to realise a single-input single-output nonlinear dynamic system simulating rainfall-runoff transformation in a urban hydrological basin. The aim is to test the performance, in simulation and real time forecasting, of these models when compared to single-input single-output linear dynamic systems with a stochastic process as forecasting component. For this reason, the impulse unit hydrograph, the transfer function of the deterministic component of such linear models, and the stochastic process have been calculated by means of the experimental data (59 events of rainfall-runoff) and, similarly, the identification procedure of the best nonlinear model was performed. The comparison between linear and nonlinear models was achieved by computing the estimated mean generalisation error and by performing statistical tests by means of cross-correlation and auto-correlation functions, using cross-validation techniques.