EFFECTS OF INPUT VARIABLE SELECTION IN ARTIFICIAL NEURAL NETWORK FOR WATER STAGE FORECASTING

Abstract

 Input variable selection is one of the most challenging tasks for modelers when building a hydrological model using artificial neural networks (ANNs). The conventional method of input variable selection for ANN considers the linear correlation of each variable with the prediction target variable. However, this conventional approach can potentially limit the ability of ANN models. This study surveys the sensitivity of input variable selection methods in ANN performance to obtain an idea to save our time and concerns related to input variable selection. We prepared three ANN models with different input variable selection methods and two regression models as well. Comparing the results from these five models, which are for hourly-based water stage prediction at the Hirakata station, indicates that ANN provides satisfactory prediction accuracy without a careful input variable selection process. And, there is a possibility that ANN performs poorly if the variable selection process eliminates the necessary data.