Estimating Total Dissolved Solids in Groundwater Using Machine Learning Models

Abstract

Accurate forecasting of water quality is pivotal for compelling pollution control and enhanced water management practices. This study predicted total dissolved solids in groundwater samples from West Bengal, India, using data sourced from the Central Pollution Control Board for the span 2020–2022. The parameters include temperature, pH, conductivity, biological oxygen demand, nitrate-N + nitrite-N, fecal coliform, total coliform, fluoride, and arsenic. Employing a diverse set of machine learning models including seven regression models, three support vector machines (SVMs), three artificial neural networks (ANNs), and an adaptive neuro-fuzzy inference system (ANFIS), the study evaluated model performance using root mean square error (RMSE), coefficient of determination (R²), mean square error (MSE), and mean absolute error (MAE). The assessment revealed that the ANN trained with Bayesian regularization emerged as the most effective, boasting the lowest errors (RMSE = 0.00147, MSE = 0.0005, MAE = 0.0112) and the highest R² (0.97), ensuring superior precision. Additionally, ANN trained with Levenberg–Marquardt and ANFIS exhibit commendable performance, showcasing minimal errors and high R² values. Among the non-ANN models, boosted tree displayed a lower RMSE (0.08246) and a higher R² (0.62), while a linear SVM demonstrated balanced performance with RMSE of 0.0877 and R² of 0.57.