Exploring the use of multi-gene genetic programming in regional models for the simulation of monthly river runoff series

Abstract

The use of new data-driven approaches based on the so-called expert systems to simulate runoff generation processes is a promising frontier that may allow for overcoming some modeling difficulties related to more complex traditional approaches. The present study highlights the potential of expert systems in creating regional hydrological models, for which they can benefit from the availability of large database. Different soft computing models for the reconstruction of the monthly natural runoff in river basins are explored, focusing on a new class of heuristic models, which is the Multi-Gene Genetic Programming (MGGP). The region under study is Sicily (Italy), where a regression based rainfall-runoff model, here used as benchmark model, was previously built starting from the analysis of a regional database relative to several gauged watersheds across the region. In the present study, different models are created using the same dataset, including: six MGGPs generated considering different modeling set-up; a Multi-Layer Perceptron Artificial Neural Network (ANN); two new hybrid models (ANN-MGGP), combining a Classifier ANN and two MGGPs that simulate separately low and high runoff. Results show how all the soft computing models perform similarly and outperform the benchmark model, demonstrating that MGGP can be considered as a valid alternative to the much more consolidated ANN technique. The new introduced hybrid ANN-MGGP is the only model showing at least satisfactory performance (i.e. Nash-Sutcliffe Efficiency above 0.5) over the full range of 38 watersheds explored, representing a useful regional tool for reconstructing monthly runoff series also at ungauged sites.