WCW26

Machine learning (ML) is increasingly applied in Water Resource Recovery Facilities (WRRFs) to improve operational decision-making, yet most implementations remain fragmented and highly dependent on either large historical datasets or narrowly defined use cases. This work presents a unified fit-for-purpose perspective on scalable ML applications for WRRFs through two complementary case studies addressing both data-rich and data-scarce operational challenges.
The first case study focuses on influent flow forecasting for proactive plant-wide optimization. Accurate forecasts enable improved chemical dosing, staffing, maintenance planning, and mitigation of wet-weather impacts in combined sewer systems. An ensemble ML framework integrating Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) models via a meta-learner was developed to leverage the usually available influent flow historical data and provide robust forecasts up to seven days ahead. Tested across 30 WRRFs (2022–2024) with average daily flows ranging from 0.3 to 146 MGD and peak flows approaching 300 MGD, the framework achieved average Mean Absolute Percentage Error (MAPE) below 11% and Nash–Sutcliffe Efficiency (NSE) around 50% for 7-day forecasts, including during extreme wet-weather events. Adaptive preprocessing tailored to dataset characteristics proved critical for consistent performance across diverse facilities.
The second case study addresses data scarcity through a new soft-sensing paradigm for monitoring biological processes without relying on historical plant data. Probabilistic ML models are trained exclusively on synthetic datasets generated from uncertainty-aware mechanistic simulations, producing priors over plausible process behavior. During deployment, predictions are refined through Bayesian updating using sparse laboratory or online measurements without model retraining. Demonstrated for real-time monitoring of both ammonia concentrations and the amount of simultaneous nitrification–denitrification (SND), the framework achieved 90% empirical coverage of true values in simulations and is currently being validated on pilot-scale using only occasional nitrogen measurements.
Together, these case studies illustrate how ML can support both predictive planning and real-time process insight across WRRFs with varying levels of data availability. By combining state-of-the-art ML methods, synthetic-data-driven probabilistic modeling, and Bayesian updating, this work outlines a scalable pathway toward uncertainty-aware digital monitoring and decision support in wastewater treatment systems.