Regional models based on Multi-Gene Genetic Programming for the simulation of monthly runoff series

Abstract

Accurate estimates of runoff in river basins are useful for several applications. The use of data-driven procedures for simulating the complex runoff generation process is a promising frontier that could allow for overcoming some typical problems related to more complex traditional approaches. This study explores soft computing based regional models for the reconstruction of monthly runoff in river basins. The region under analysis is the Sicily (Italy), where a regressive rainfall-runoff model, here used as benchmark model, was previously built using data from almost a hundred gauged watersheds across the region. This previous model predicts monthly river runoff based on a unique regional, non-linear, regression equation with four site-specific parameters, where regressors are precipitation, temperature and runoff at the previous month; the site-specific parameters are assessed using a further set of equations that account for some physical attributes of the basins. In this study, the underlying relationship between monthly runoff and the same predictors is explored without any a-priori assumption about its analytical form, creating different types of soft computing regional models, i.e. different Multi-Gene Genetic Programming (MGGP) models, generated considering two different set-up (simple and complex) and three different sets of input variables, and an Artificial Neural Network (ANN). Results highlight the potential of data-driven expert systems in creating regional hydrological models, for which they can benefit from the availability of large database; the comparison among the models shows, in fact, how both the MGGP and the ANN models outperform the benchmark model.