Real time forecasting of runoff in a urban hydrological catchment: comparison among ANN, SVM and GP

摘要

Hydrological and hydraulic processes determine inflow-outflow transformation phenomenon in catchment areas in a complex way. An interesting issue in technical hydrology, i.e. urban drainage networks is to predict runoff due to rainfall. Modeling this kind of phenomena is a very interesting aim because the behavior of such a physical system is neither linear nor time invariant. This means that a single Unit Hydrograph is not able to describe the hydrological processes during all the possible events of rainfall. In the last years, many authors modeled hydrological systems by means of Artificial Neural Networks, Support Vector Machines and Genetic Programming in order to remove liner hypothesis of models based on Unit Hydrography. These techniques are called data-driven because they allow to use monitoring data to perform real time forecasting of runoff (on-line prediction) and off-line prediction (simulation). In this work, Artificial Neural Networks, Support Vector Machines and Genetic Programming data-driven techniques have been critically compared from a theoretical point of view. Then aim is to test their performances especially in real time forecasting. For this reason, models built on basis of Artificial Neural Networks, Support Vector Machines and Genetic Programming techniques have been realized by means of the experimental data from the urban watchment of Luzzi, in South Italy, that is a well monitored small watchment are abuseful in testing data-driven rainfall-runoff modeling. The study of the performance of these non-linear moles has been achieved by computing the estimated mean generalization error of k-step-ahead predictions, using cross-validation.