Modelling point-of-consumption residual chlorine in humanitarian response: Can cost-sensitive learning improve probabilistic forecasts?

Ensuring sufficient free residual chlorine (FRC) up to the time and place water is consumed in refugee settlements essential for preventing spread of waterborne illnesses. Water system operators need accurate forecasts FRC during household storage period. However, factors that drive decay after leaves piped distribution vary substantially, introducing significant uncertainty when modelling point-of-consumption FRC. Artificial neural network (ANN) ensemble forecasting systems (EFS) can account this by generating probabilistic ANNs are typically trained using symmetrical error metrics like mean squared (MSE), but leads forecast underdispersion (the smaller than observations). This study proposes solve training an ANN-EFS cost functions combine alternative (Nash-Sutcliffe efficiency, Kling Gupta Efficiency, Index Agreement) with cost-sensitive learning (inverse weighting, class-based inverse frequency weighting). The each function was evaluated quality data from Bangladesh Tanzania comparing percent capture, confidence interval reliability diagrams, rank histograms, continuous ranked probability. Training developed produced a 70% improvement dispersion compared baseline best performance obtained model Kling-Gupta Efficiency weighting. Our findings demonstrate improve better post-distribution decay. These techniques enable humanitarian responders ensure more reliably at point-of-consumption, thereby