Continual Learning with Elastic Weight Consolidation for Robust Combined Sewer Overflow Optimization via Neural Inversion
Conference
Syed, Z, Saadati, Y, Yin, Z et al. (2026). Continual Learning with Elastic Weight Consolidation for Robust Combined Sewer Overflow Optimization via Neural Inversion
. 548-559. 10.1061/9780784486931.045
Syed, Z, Saadati, Y, Yin, Z et al. (2026). Continual Learning with Elastic Weight Consolidation for Robust Combined Sewer Overflow Optimization via Neural Inversion
. 548-559. 10.1061/9780784486931.045
Traditional supervised Artificial Intelligence (AI) methods struggle with unseen dynamic conditions and are vulnerable to catastrophic forgetting (CF), which requires computationally expensive full retraining. To enable real-time adaptability in urban systems, this study presents a continual learning framework using Elastic Weight Consolidation (EWC) to preserve prior knowledge while integrating new data. The framework’s efficacy was tested in a Combined Sewer Overflow (CSO) optimization environment via a novel Neural Network Inversion technique, using baseline and four sequential task datasets that represent evolving precipitation patterns. While EWC alone effectively learned new tasks (achieving R2 0.90), it exhibited severe CF on the baseline task, where the R2 plummeted from 0.96 to −28.81. Incorporating a replay buffer fully mitigated this issue, successfully maintaining baseline task performance (R2 = 0.93) and significantly reducing the RMSE of CSO volume prediction from 4,422.0 m3 to 216.7 m3. The EWC with Replay Buffer framework on all tasks gave the lowest error during optimization. This scalable and computationally efficient approach improves optimal control decisions and leads to better CSO reduction, offering a valuable solution for urban systems facing rapid environmental changes.
