Flexible recoveries of water treatment plants from anomalies through self-learning from digital twin-based simulations

PI: Pingbo Tang

Co-PI(s): Yorie Nakahira

University: Carnegie Mellon University

Industry partners: Evoqua Water Technologies, Ethos Collaborative, Pennsylvania American Water

The proposed PITA project aims to enable flexible recoveries of water treatment plants from anomalies in their disinfection and filtration processes to reduce operating costs while ensuring safety and public health. Water treatment plants need disinfection control and filter backwashes to meet water quality standards while ensuring safety and cost-effectiveness. Anomalies in disinfection and filtration processes are deviations from expected process efficiency, treatment costs, toxic byproducts, and other water quality criteria. System faults or changed water influents or demands trigger these issues. State-of-the-art process control algorithms need operators in the control loop to handle unexpected scenarios. Occasionally, improper anomaly handling causes violations of water quality regulations, expensive chemical treatments, machine faults, and high operating costs. The project team proposes a self-learning framework that allows computers to synthesize reusable anomaly recovery strategies from many plants' operation histories and digital twins reconstructed from these histories. Existing water system control methods could hardly use limited historical data to learn comprehensive anomaly recovery strategies. We will 1) use limited histories for developing data-driven simulations of anomaly handling scenarios in three plants and 2) establish an inverse reinforcement learning approach that identifies reusable recovery strategies from simulations of handling four types of disinfection/filtration anomalies.