Integrated Continual Learning Frameworks for Real-Time Combined Sewer Overflow Control Prediction and Optimization via Neural Inversion

These data form part of a research submission and primarily contain the outputs generated by the conducted study. The research presents an integrated framework for managing Combined Sewer Overflows (CSOs) in Detroit’s Puritan–Fenkell/Seven-Mile Collection System (PFSMC). It addresses the challenge of climate non-stationarity, wherein evolving precipitation patterns reduce the reliability of traditional stationary models. The study employs Continual Learning (CL) strategies specifically regularization-based, memory-based, and Bayesian approaches to incrementally update deep neural network (DNN) surrogate models without requiring full retraining. This predictive framework is coupled with neural inversion to optimize real-time control (RTC) of sewer infrastructure, such as gates and pumps, with the objective of maximizing storage utilization and reducing overflow events.