Please login first
Optimized chlorine bulk decay models and a machine-learning-guided water quality responsive kinetic model for residual chlorine prediction
1 , 2 , 3 , 4 , * 5
1  PhD candidate at the National University of Singapore
2  Senior development engineer and project manager at Xylem Water Solutions Singapore
3  Senior research fellow at the National University of Singapore
4  Associate Professor at the National University of Singapore
5  Professor at the National University of Singapore
Academic Editor: Lampros Vasiliades


Effective treatment and disinfection of source waters and the safe delivery of potable water to customers require a comprehensive understanding of water quality changes from source to tap. Variations in source water quality, treatment and supply operations, and water demand present significant challenges to maintaining consistent water quality across the system and can lead to areas that suffer from low residual, nitrification episodes, and chemical or microbial water quality violations. Prediction of water quality within distribution systems, including disinfectant residual loss and by-product generation, has been a subject of applied research since the early 1990’s. Almost all scientists who have suggested existing models of chlorine decay in the literature state that their models rank among the most effective models of chlorine decay. However, there have been numerous disputes and discussions among researchers and experts regarding the superiority of certain models over others. In this study the currently existing process based bulk decay models were modified by replacing initial chlorine concentration parameter with chlorine demand in their equations and the results showed that this modification could improve the performance of the models by 38.03%, 28.02%, 23.11% and 33.29% in average for First Order Model (FOM), Parallel First Order Model (PFOM), Second Order Model (SOM) and Parallel Second Order Model (PSOM), respectively. In addition, it was proven that the chlorine decay prediction in water distribution system can be modified and robust to be used as an online tool for predicting residual chlorine in different locations of distribution system over time rather than to be restricted by off-line use and planning-level analysis. In this regard, an online predictive method based on a machine learning algorithm was introduced and implemented in this study to predict first order chlorine bulk decay rate by feeding water quality parameters as the inputs. Hence, a Gaussian Process Regression (GPR) model was trained and used to predict the kinetic parameter in FOM, and the results showed that although the accuracy of predictions for the test set was high for most of the cases, the high sensitivity of the FOM to its kinetic parameter (first order decay rate coefficient) resulted in high MSE values in some of the Total Residual Chlorine (TRC) predictions. However, the high correlation coefficients between the predicted and actual TRC values represents the fact that the model could properly identify the substantial process behind TRC prediction based on water quality parameters. In addition, a novel methodology was introduced and suggested in this study based on the obtained results to be applied in real water distribution system for an optimized online prediction of residual chlorine. By incorporating the variability of source natural organic matter, along with operational actions and water demands, the proposed approach seeks to address a long-standing research challenge to develop high fidelity and robust water quality predictions – well suited to providing operational decision support for optimized distribution system management. This research is supported by the National Research Foundation, Singapore, and PUB, Singapore’s National Water Agency under its Urban Solutions & Sustainability Competitive Research Program (Water) PUB-1804-0084.

Keywords: Chlorine decay; Distribution system; Bulk water decay; Water quality; Modelling; Machine Learning