Predicting Water Quality Indicators from Conventional and Nonconventional Water Resources in Algeria Country: Adaptive Neuro-Fuzzy Inference Systems Versus Artificial Neural Networks

Abstract

Monitoring water quality is of great importance and mainly adopted for water pollution control of conventional and nonconventional water resources. Generally, water quality is evaluated using several indicators, including chemical oxygen demand (COD), biochemical oxygen demand (BOD), and dissolved oxygen concentration (DO). In the present investigation, two artificial intelligence techniques, namely, adaptive neuro-fuzzy inference system (ANFIS) and artificial neural networks (ANN), were applied for predicting two water quality indicators: (1) chemical oxygen demand (COD) at Sidi Marouane Wastewater Treatment Plant (WWTP), east of Algeria, and (2) dissolved oxygen concentration (DO) at the drinking water treatment plant of Boudouaou, Algeria. The models were developed and compared based on several water quality variables as inputs. Three ANFIS models, namely, (1) ANFIS with fuzzy c-mean clustering (FCM) algorithm called ANFIS_FC, (2) ANFIS with grid partition (GP) method called ANFIS_GP, and (3) ANFIS with subtractive clustering (SC) called ANFIS_SC, were developed. The ANFIS models were compared to standard multilayer perceptron neural network (MLPNN) and multiple linear regression model (MLR). Results obtained demonstrated that (1) for predicting COD, ANFIS_SC is the best model, and the coefficient of correlation (R), Wilmot’s index (d), root-mean-square error (RMSE), and mean absolute error (MAE) were calculated as 0.805, 0.880, 6.742, and 4.944 mg/L for the validation dataset. The worst results were obtained using the MLR model with R, d, RMSE, and MAE equal to 0.750, 0.840, 0.7658, and 5.916 mg/L for the validation subset, and (2) for predicting DO concentration, the best results were obtained using ANFIS_SC with R, d, RMSE, and MAE equal to 0.856, 0.922, 1.528, and 1.123 mg/L for the validation subset, respectively.