A data-driven model for the prediction of chlorine losses in water distribution trunk mains

Abstract

Abstract A data-driven model that uses 4 different machine learning (ML) algorithms (Feed forward artificial neural networks (ANN), Nonlinear autoregressive exogeneous (NARX) ANN, support vector machine and Random Forest) was designed for the prediction of chlorine loss events in water distribution trunk mains. The model, firstly, identifies past chlorine loss events and their associate flow or temperature events. Then, the detected past flow events and their associate past chlorine loss events are used to train the ML algorithms. The model was tested in 3 trunk mains of the same drinking water distribution system with similar diameter but with different characteristics, using each time a different combination of parameters (flow (input) - past chlorine losses (output) or flow, temperature, and chlorine (input) - past chlorine losses (output)) and machine learning algorithms. Results indicate that the model could predict a future chlorine loss event with a period between 2 to 10 hours depending on the parameter and ML algorithm used and the trunk mains’ hydraulic characteristics.