Data-driven rainfall/runoff modelling based on a neuro-fuzzy inference system

The development of rainfall/runoff models involves extensive computation and the availability of different coexisting platforms, including numerical flow models and GIS for their physiographical characterization. In this paper we present a data-driven approach which avoids the use of GIS, but is based on a combination of Principal Component Analysis (PCA) and an Adaptive Neuro Fuzzy Inference System (ANFIS) to produce a simple and effective output flow prediction based on previous rainfall/ runoff data in the catchment. The emphasis of the paper is on how to set-up an efficient data structure that produces a good output flow estimation. The PCA approach is compared to the Thiessen polygons method, requiring GIS, and we demonstrate that the former can produce a better ANFIS model, with less algorithmic complexity and improved accuracy. The combined PCA þ ANFIS procedure is applied to two minor river basins in Tuscany, Italy, to demonstrate its effectiveness.