Explainable artificial intelligence for chemical dosing in drinking water treatment: A SHAP-assisted case study.

Abstract

Determining the optimal chemical dosing in drinking water treatment plants is a complex process due to the variability of hydro-meteorological parameters and the nonlinear dynamics of water chemistry. Conventional machine learning approaches typically rely on a behavioral cloning paradigm, thereby replicating suboptimal or inefficient operator decisions. In this study, a novel strategy is proposed to transform artificial intelligence from a purely predictive tool into a prescriptive decision support system. World Health Organization (WHO) and Turkish Standards Institute (TS-266) water quality standards are applied as a pre-filter, and the models are trained exclusively on data from optimal operational conditions. Tree-based ensemble algorithms are employed to predict the dosing of pre-chlorination, post-chlorination, ferric chloride, and potassium permanganate. To improve model transparency, the explainable artificial intelligence (XAI) method SHAP (SHapley Additive exPlanations) is integrated. Rather than establishing physical causality, SHAP is utilized to identify the most influential variables contributing to the model's predictions. Consequently, SHAP dependence plots are used to derive actionable, site-specific operational rules for plant operators. While the proposed framework demonstrates high predictive accuracy, the current study is limited to a single-site dataset, highlighting the need for future external validation to fully assess model generalizability and associated uncertainties. Ultimately, this framework integrates explainable machine learning with operational principles, offering a reliable foundation to reduce chemical consumption and minimize human error.