Evaluation of Process-Based Ensemble Models for Forecasting Point-of-Consumption Free Residual Chlorine in Refugee Settlements

Waterborne illnesses are a leading public health concern in refugee and internally displaced person (IDP) settlements. Controlling the spread of these illnesses can be particularly challenging as pathogens can be introduced into previously-safe drinking water during the post-distribution period of collection, transport, and household storage. Free residual chlorine (FRC) is often used in these settlements to prevent recontamination of drinking water, and thus, it is critical that at least 0.2 mg/L of FRC is available up to the point-of-consumption. Chlorine decay models can be used to determine the chlorine dose required to maintain this residual; however, post-distribution FRC decay is highly uncertain due to many immeasurable factors that vary substantially from user to user within a site. Traditional deterministic FRC decay models are unable to quantify this uncertainty. Therefore, there is a need for improved modelling that quantifies uncertainty in FRC decay. Ensemble forecasting systems, which consist of collections of models as opposed to a single standalone model, can quantify this uncertainty by generating probabilistic forecasts of FRC. This study presents a novel use of ensemble techniques to generate probabilistic forecasts of FRC decay for the post-distribution period in refugee and IDP settlements. The two alternatives considered for determining the decay parameters for the ensembles were a resampling approach with least-squares regression and a quantile regression-based approach, both using six different FRC decay equations. These approaches were tested using a six-month operational water quality dataset collected from a refugee settlement in Bangladesh in 2019. The quantile regression-based ensembles produced more reliable forecasts, and better capture of observed values, as compared to resampling with least-squares. Of the FRC decay equations considered, the parallel first-order decay equation produced the least quantile error when compared to the other decay equations considered. This demonstrates that ensemble forecasting systems effectively quantify uncertainty when modelling post-distribution FRC decay. These findings can be used to develop improved FRC guidance for humanitarian responders working in refugee and IDP settlements.