Surface Water Quality Model: Impacts of Influential Variables

Abstract

Considering all possible input (predictor) variables in a predictive water quality model is impractical owing to computational workload and complexity of the problem. Computational efficiency, as well as complexity of a model, is greatly increased if the most influential variables are determined using an input variable selection technique. In this study, the multilayer perceptron artificial neural network was implemented in order to predict total dissolved solids in the Sufi Chai river (Iran). The research studied the impacts of chemical composition (salinity, potassium, sodium, magnesium, calcium, sulfate, chloride, bicarbonate, carbonate, pH, and sodium adsorption ratio) in source water, climatic variables (rainfall, air temperature, wind speed, and evaporation), and hydrometric variables (river discharge and suspended sediment) on the predictions. Garson’s equation was used to find the relative importance of each input variable and to select the most influential variables. A correlation method was applied and the results were compared with those of the Garson method. A set of 12-year data (1999–21) was used to calibrate, validate, and test the models. The results indicated that input variable selection before modeling can improve both accuracy and simplicity of the models. Although Garson and correlation methods both improved the accuracy of the models, the Garson method was found to be more accurate. As well, the research showed including climatic and hydrologic variables improved the accuracy of the models with fewer variables considered.