Prediction of electrical conductivity using ANN and MLR: a case study from Turkey

Abstract

The study areas are located in Turkey (Kastamonu, Bartin, Karabuk, Sivas) and contain very different rock types, various mining and agricultural activity opportunities. So, the areas have groundwaters that have different chemical compositions and electrical conductivity (EC) values. The EC can be measured using EC meter, and it must be measured in situ. But, the measurement of EC in situ is laborious, time-consuming, expensive, and difficult in arduous terrain environments. In recent years, machine learning models have been a primary focus of interest for a lot of study by providing often highly accurate forecast for solutions of such problems. The aim of the study is to forecast EC of groundwater using artificial neural networks (ANN) and multiple linear regressions (MLR). Twelve different hydrochemical parameters, which affect the EC, such as major/minor ions and trace elements, were used in the analysis. Multilayer feed-forward ANN trained with backpropagation in Python machine learning libraries was used in this study. In order to obtain the most appropriate ANN architecture, trial-and-error procedure was used and different numbers of hidden layers, neurons, activation functions, optimizers, and test sizes were constructed. This study also tests the usability of input parameters in EC prediction studies. As a result, comparisons between the measured and predicted values indicated that the machine learning models could be successfully applied and provide high accuracy and reliability for EC and similar parameters forecasting.