Comparison of Electrical Conductivity Prediction Models Using Gaussian Process

Abstract

People living in coastal areas use clean water sourced from groundwater to support the household, agricultural, and industrial needs. However, human activities and natural factors can lead to a common problem in coastal areas, namely seawater intrusion. Seawater intrusion can be detected using water quality data. Today, one of the challenges in water resources management is the prediction of water quality parameters such as total dissolved solids (TDS), electrical conductivity (EC), and water turbidity. Incomplete EC data and limitations of direct measurements can affect the analysis. Machine learning models are known to provide the most accurate predictions. This research used EC parameter data to investigate the performance of algorithms, namely artificial neural networks (ANN), Gaussian processes (GP), and multiple regression (MLR). The prediction used seven hydrochemical parameters (K, Ca, Mg, Na, SO4, Cl, HCO3) and three physical parameters of groundwater (TDS, pH, EC). Performance measurement used R-squared (R2) and root mean squared error (RMSE). The testing showed the MLR model had R2 of 0.985 and RMSE of 0.030, which were slightly better than other models. Hence, it can be concluded that the MLR model can be a solution to difficult problems of EC prediction and incomplete data in the water resources management.