Water Supply (Mar 2022)
Integrating water quality and streamflow into prediction of chemical dosage in a drinking water treatment plant using machine learning algorithms
Abstract
Understanding the relationship between raw water quality and chemical dosage is especially important for drinking water treatment plants (DWTP) that have multiple water sources where the ratio of different supply sources could change with seasons or in a matter of weeks in response to changing hydrologic conditions. In this study, the potential for deploying machine learning algorithms, including principal component regression (PCR), support vector regression (SVR) and long short-term memory (LSTM) neural network, are tested to build predictive models. These tools were used to estimate chemical dosage at a daily time-scale. Influent water quality such as pH, color, turbidity, and alkalinity, as well as chemical dosage including sulfuric acid, ferric sulfate and liquid oxygen were used to build and test these models. An 80/20 percent data split was used for training and testing model performance using correlation coefficients, relative mean square error, relative root mean square error and Nash–Sutcliffe efficiency. Results indicate, compared with PCR, both SVR and LSTM were able to capture the nonlinear relationship between chemical dose and source water quality changes and displayed higher predictive skills. These types of models have application in real-time operational support without requiring computationally expensive physics-based models. HIGHLIGHTS A study that examines water quality and chemical use from a water treatment plant.; Appropriate chemical dosage is essential to ensure safe potable water.; Understanding the relationship between water quality and chemical use is a key step.; Different algorithms are tested to build predictive models.; Support vector machine and neural networks better capture nonlinear relationships.;
Keywords
