Improving generalization of artificial neural networks in rainfall-runoff modelling

Abstract

Artificial neural networks (ANNs) are general-purpose techniques that can be used for nonlinear data-driven rainfall-runoff modelling. The key issue to construct a good model by means of ANNs is to understand their structural features and the problems related to their construction. Indeed, the quantity and quality of data, the type of noise and the mathematical properties of the algorithm for estimating the usual large number of parameters (weights) are crucial for the generalization performances of ANNs. However, it is well known that ANNs may suffer from poor generalization properties due to the high number of parameters and non-Gaussian data noise. Therefore, in the first part of this paper, the features and problems of ANNs are discussed. Eight Avoiding Overfitting Techniques are then presented, considering that these are methods for improving the generalization of ANNs. For this reason, they have been tested on two case studies-rainfall-runoff data from two drainage basins in the south of Italy-in order to gain insight into their properties and to investigate if there is one that absolutely gives the best performance.