Estimation of total dissolved solids (TDS) using new hybrid machine learning models

Abstract

The overall quality of Groundwater (GW) is important, primarily because it determines the suitability of water for drinking, irrigation, and domestic purposes. In this study, the adaptive fuzzy interface system (ANFIS), support vector machines (SVMs), and artificial neural network (ANN) models were employed for predicting the total dissolved solids of aquifers. The moth flam optimization, cat swarm optimization (CSO), particle swarm optimization (PSO), shark algorithm (SA), grey wolf optimization (GWO), and gravitational search algorithm (GSA) were used to train the ANFIS, SVM, and ANN models. The data were collected from Yazd plain (Iran) to predict the Total Dissolved Solids (TDS). The principal component analysis was used to determine the most appropriate inputs for predicting TDS. The hybrid ANFIS-MFO improved the accuracy of RMSE (roo mean square error) over the ANN-MFO and SVM-MFO models by 1.4% and 3.8%, respectively. It was also observed that the SVM model had the least NSE (Nash Sutcliffe efficiency) value among all the models. Unlike the standalone ANFIS, the multilayer perceptron (MLP), and SVMs models, the hybrid ANFIS, ANN, and SVM demonstrated high accuracy in the training and testing phase, so that in the optimal hybrid model, ANFIS-MFO, values of mean absolute error (MAE), Nash Sutcliff efficiency (NSE), and percent bias (PBIAS) were 2.21 (mg/lit), 0.94, 0.15, 2.981 (mg/lit), 0.93, and 0.18, respectively. The ANFIS-MFO was also seen to further enhance the RMSE by approximately 3% and 7%, as compared to the ANN-MFO and SVM-MFO. This study also aims to investigate the temporal variability TDS using innovative trend analysis (ITA). The TDS value of < 1800 (mg/lit) indicates a decreasing trend, while a medium TDS value (2000 mg/lit < TDS < 2800 mg/lit) does not have a significant trend. The high TDS values (TDS > 3000 mg/lit) indicate an increasing trend. In this study, the ANFIS-MFO and ANFIS-CSO models showed superior performance over the other models; hence, indicates significant implication in their application for other water resources and hydrological variables.