Space-time prediction of residual chlorine in a water distribution network using artificial intelligence and the EPANET hydraulic model

Insufficient knowledge of physical models and difficulty in fitting statistical models impair the choice of models to regulate residual chlorine in water distribution. This paper compared the performance of physical and statistical models in predicting residual chlorine concentrations in drinking water distribution. Drinking water was sampled from the downstream 128 water points water pipeline. Online chlorine concentrations were determined at water draw-off points. EPANET, the physical model, was used because of its efficiency in tracking dissolved chemicals. Statistical models used were regression, decision tree, random forest and artificial neural network. In the whole distribution network, the artificial neural network performed at R-2 of 94%, multi-linear regression (62%), random forest (55%), decision tree (41%), and EPANET (24%). However, EPANET yielded improved performance with R-2 above 70% when separately applied to individual sub-distribution networks; hence, is recommended for secondary chlorination in small distribution networks. For modelling large distribution networks, statistical models, especially an artificial neural network, are recommended. However, such cases still need support from confirmatory systems of interpretable parametric or hydraulic models that can achieve good performance with R-2 >= 80%. Water utilities can use these results to deploy model(s) for managing residual chlorine within safe limits of residual chlorine concentration in water distribution practice.