Physics-informed ensemble machine learning for dynamic water quality prediction in water distribution systems

Maintaining water quality (WQ) across water distribution systems (WDSs) remains a persistent challenge, as it is governed by complex interactions involving multiple chemical and biological species. Physics-based models (PBMs) have been traditionally used as the primary approach for predicting WQ dynamics in WDSs. In PBMs, the WQ is simulated by numerically solving the fundamental partial differential equations (PDEs) that govern the fate and transport of different species in the pipes of WDSs. These models are computationally intensive, which restricts their use in real-time control applications, particularly for large and complex WDSs. Recently, machine learning (ML)-driven WQ prediction frameworks have gained significant traction as an alternative to PBMs due to their flexibility and computational efficiency. However, unlike PBMs, ML models are typically designed and trained for a specific WDS, limiting their applicability to other systems. This study seeks to address this key limitation by introducing novel Physics-Informed Machine Learning (PI-ML) approaches for simulating WQ in WDSs. The proposed framework involves training and integrating an ensemble of ML models to capture the key transport and reaction processes that make up the governing PDEs of PBMs, including advection, dispersion, and reaction. The results show that integrating fundamental physical insights into the development of ML algorithms improves predictive accuracy and generalizability compared to data-driven ML models. • Developed a physics-informed machine learning framework for water quality in distribution systems. • Incorporated advection, dispersion, and reaction into ensemble machine learning models. • Integrating physical insights improved accuracy and generalizability of machine learning models. • Bayesian model averaging of base learners improved performance and reduced overfitting.