Arsenic spike prediction in a Hungarian drinking water distribution network using interpretable machine learning

Arsenic is a global contaminant of concern as chronic exposure above 10 μg/L in drinking water results in health issues including cancer, impaired cognitive development, and skin lesions. Arsenic is naturally abundant in Hungarian groundwater, posing challenges for compliance with European Union regulations. Trace amounts of arsenic accumulate in water distribution system pipes and may be mobilized downstream by fluctuations in pressure and water quality. Additionally, ambient arsenic can infiltrate from surrounding groundwater into compromised pipes during low pressure conditions. This research identifies mechanisms of interaction between water quality and pressure parameters, based on samples collected at ten public water faucets in an at-risk system on the Hungarian Great Plain, to reliably predict transient, localized arsenic spikes. Arsenic species were separated using an ion-exchange resin and measured via inductively coupled plasma mass spectrometry (ICP-MS) or inductively coupled plasma optical emission spectroscopy with hydride generation (ICP-OES). Machine learning models predicted arsenic concentration using pressure data from nine hydrant-mounted sensors and water quality measurements. Random forest modeling achieved an excellent F1 score of 0.866 on unseen data. Manganese concentration and manganese interactions with iron, temperature, and mean and maximum pressure severity indices were the most important features aside from finished water arsenic concentration (SHAP = 0.229, 0.126, 0.095, 0.061, respectively). pH, redox potential, and temperature interactions with pressure events constituted the majority of the remaining features. Results were robust to exclusion of finished water arsenic concentration and missing data imputation strategy. Findings indicate that complex water quality and pressure interactions mobilize arsenic within the distribution network, yielding important implications for system management.