Modified Relative Strength Effect to Facilitate Artificial Neural Network Development for Hydrologic Data

Abstract

A new index called a modified relative strength effect (RSE) is developed for determining the influence of an input in an artificial neural network (ANN) model. This modified RSE, which is an improvement over the usual RSE, indicates the influence of each input on the target output in different ranges and can be estimated easily during ANN model development by a user. The potential of this modified RSE was examined using ANN models developed for finding atypical bacteria counts and daily flow. The modified RSE and traditional RSE were used to identify the essential inputs for the ANN models in this research. The results show an improvement of 10% in the mean square error when the modified RSE is used for input selection instead of the traditional RSE. This study also explores the usefulness of the modified RSE in identifying the required input lags while developing an ANN time-series model. A time-series model was constructed using ANN, and the modified RSE values of inputs were compared to the partial autocorrelation coefficients (PCCs) of the first 12 lags. The PCCs were calculated using standard procedure. Perfect matching of the modified RSE with the PCC was observed. The results indicate that the RSE can be used like the PCC while developing ANN-based time-series models.