115 - Physics-Informed Stacking Ensemble Machine Learning for Water Quality State Estimation in Water Distribution Systems

Disinfectant chemicals play a crucial role in maintaining drinking water quality and safety. Chlorine is the most widely used disinfectant in water distribution systems (WDSs). Physically based models (PBMs) have typically been the primary approach for understanding and predicting the transport and decay of chlorine in WDSs. In PBMs, the fundamental partial differential equations controlling the complex, dynamic behavior of chlorine within the network are numerically solved. Hence, these models are computationally expensive, which limits their applications for real-time simulation, especially for large WDSs. In recent years, machine learning (ML) models have emerged as a promising alternative to PBMs. ML models offer the advantage of flexibility and adaptability, making them well-suited for complex systems like WDSs. However, unlike PBMs, data-driven ML models are typically developed and trained for a specific WDS and hence have limited generalizability to other systems. This study aims to overcome this limitation by proposing a novel physics-informed stacking ensemble machine learning approach for water quality state estimation in WDSs. The approach will implement stacking techniques to leverage ensemble behaviors by combining the outputs of multiple individually trained ML models. Each individual model will be trained using analytical and numerical evaluations of each of the fundamental transport processes that control chlorine fate and transport in the WDS pipes, namely, advection, dispersion, and reaction. The integration and compilation of various physical processes within the stacked ensemble ML algorithm will help enhance its generalizability and improve its predictive accuracy compared to purely data-driven ML approaches.