Chemical Dosage Prediction for Drinking Water Treatment Using Random Forest and Polynomial Regression

sally elrashedy , dingbao wang, Tirusew Asefa, Jack Thornburgh, Hui Wang Hui Wang

Predictive modeling of chemical dosage based on raw water quality can be useful in decision-making in operating a water treatment plant. In this study, two statistical methods, i.e., random forest and polynomial regression, are used for modeling the chemical usages in drinking water treatment based on the measured water quality parameters in source water. The daily chemical dosages and eight water quality parameters were recorded from 2014 to 2021. The dataset is split into training (80%) and testing (20%) sets. For the training set, random forest has better performance than polynomial regression for all three chemical usages. However, random forest and polynomial regression have comparable performance for the testing set; particularly, the performance of polynomial regression on modelling sulfuric acid is slightly better than that of random forest, but random forest has a slightly better performance on modelling ferric sulfate and Actiflo polymer. The polynomial regression model is robust due to its explicit polynomial equation and its interaction terms capturing the impact of the interdependency of water quality parameters. It is found that both random forest and polynomial regression are able to model the dosages of sulfuric acid and ferric sulfate effectively, and the dosage of Actiflo polymer acceptably.