Modeling Water-Quality Parameters Using Genetic Algorithm–Least Squares Support Vector Regression and Genetic Programming

The modeling and monitoring of water-quality parameters is necessary because of the ever increasing use of water resources and contamination caused by sewage disposal. This study employs two data-driven methods for modeling water-quality parameters. The methods are the least-squares support vector regression (LSSVR) and genetic programming (GP). Model inputs to the LSSVR algorithm and GP were determined using principal component analysis (PCA). The coefficients of the LSSVR were selected by sensitivity analysis employing statistical criteria. The results of the sensitivity analysis of the LSSVR showed that its accuracy depends strongly on the values of its coefficients. The value of the Nash-Sutcliffe (NS) statistic was negative for 60% of the combinations of coefficients applied in the sensitivity analysis. That is, using the mean of a time series would produce a more accurate estimate of water-quality parameters than the LSSVR method in 60% of the combinations of parameters tried. The genetic algorithm (GA) was combined with LSSVR to produce the GA-LSSVR algorithm with which to achieve improved accuracy in modeling water-quality parameters. The GA-LSSVR algorithm and the GP method were employed in modeling Na+, K+, Mg2+, SO42−, Cl-, pH, electric conductivity (EC), and total dissolved solids (TDS) in the Sefidrood River, Iran. The results indicate that the GA-LSSVR algorithm has better accuracy for modeling water-quality parameters than GP judged by the coefficient of determination (R2) and the NS criterion. The NS static established, however, that the GA-LSSVR and GP methods have the capacity to model water-quality parameters accurately.