A comparative study of prediction models for Coagulation process in drinking Water Treatment Plants

Coagulation process is an essential part of drinking water treatment operations. The jar test is a widely-used laboratory procedure to determine the quantity of coagulation chemical dosages for water treatment plants. However, the test is not effective for real-time control of this process due to the nonlinear behaviour and rapid variations of water quality parameters. In this study, empirical models to predict the coagulation chemical dosages at a water treatment plant in South Africa using adaptive neurofuzzy inference system (ANFIS), artificial neural network (ANN), and genetic programming (GP) are developed and compared. Ten water quality parameters are used as the model inputs while the primary coagulant, secondary coagulant, and pH adjustment chemical dosages serve as the three model outputs. The performances of ANFIS, ANN and GP models are evaluated using root mean square error (RMSE). Simulation results show that ANFIS model has the lowest RMSE values and best estimation capability when compared with the ANN and GP models. However, the GP model may provide insight into influencing input variables on the process.