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

Published: 2 January 2026| Version 1 | DOI: 10.17632/4spjvszr6z.1
Contributor:
Zain Syed

Description

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.

Files

Steps to reproduce

To reproduce the study's results objectively, follow these four steps: Establish Environment and Data: Build a hydraulic simulation of Detroit’s PFSMC system using EPA SWMM and PySWMM. Generate 24,000 baseline samples from historical NOAA records and 12,000 samples per task (T1–T4) using bias-corrected NA-CORDEX future climate projections . Implement Neural Architectures: Develop Standard DNNs (SDNN) for fixed-weight modeling and Bayesian DNNs (BDNN) for probability-based weight distributions . These form the surrogate models used to predict total CSO volumes based on precipitation and control actions. Deploy Continual Learning (CL): Introduce Tasks 1–4 sequentially to the model, utilizing VCL, EWC, or EVCL strategies. Critically, integrate a Replay Buffer to store and periodically re-train on past data samples, which effectively mitigates "catastrophic forgetting" of the baseline system behavior . Execute Optimization and Validation: Use neural inversion to minimize overflows by backpropagating loss gradients directly to the input control actions (gates and pumps) . Validate these optimized actions by running them through the SWMM simulation to measure actual reduction in CSO and efficiency in storage utilization.

Institutions

  • Florida International University

Categories

Urban Water, Urban Water Management

Funders

Licence