Input Selection of Wavelet-Coupled Neural Network Models for Rainfall-Runoff Modelling

Abstract

The use of wavelet-coupled data-driven models is increasing in the field of hydrological modelling. However, wavelet-coupled artificial neural network (ANN) models inherit the disadvantages of containing more complex structure and enhanced simulation time as a result of use of increased multiple input sub-series obtained by the wavelet transformation (WT). So, the identification of dominant wavelet sub-series containing significant information regarding the hydrological system and subsequent use of those dominant sub-series only as input is crucial for the development of wavelet-coupled ANN models. This study is therefore conducted to evaluate various approaches for selection of dominant wavelet sub-series and their effect on other critical issues of suitable wavelet function, decomposition level and input vector for the development of wavelet-coupled rainfall-runoff models. Four different approaches to identify dominant wavelet sub-series, ten different wavelet functions, nine decomposition levels, and five different input vectors are considered in the present study. Out of four tested approaches, the study advocates the use of relative weight analysis (RWA) for the selection of dominant input wavelet sub-series in the development of wavelet-coupled models. The db8 and the dmey (Discrete approximation of Meyer) wavelet functions at level nine were found to provide the best performance with the RWA approach.