Robustness of neuro-evolution in urban drainage system control under communication failures: comparing centralized and decentralized control schemes

Abstract

Real-time control (RTC) in urban drainage systems can effectively mitigate flooding and Combined Sewer Overflow (CSO) spills. Recently neuro-evolution has shown promise in RTC, which relies on communication systems to receive real-time state information and send control signals. However, the impact of communication system failures on this approach is not fully understood. This study aims to evaluate the robustness of neuro-evolution for urban drainage system operation under various communication failure scenarios, focusing on both centralized and decentralized control schemes. The communication failures considered in this study include transient disruptions in the observation or action communication process and prolonged sensor disconnections. The simulation results from the Astlingen benchmarking network indicate that the performance in total CSO volume reduction ranks as follows: centralized neuro-evolution > decentralized neuro-evolution > the baseline strategy: Equal Filling Degree (EFD). In terms of robustness, centralized neuro-evolution outperforms under observation communication disruptions and sensor disconnections, while decentralized neuro-evolution excels in handling action communication disruptions and maintaining local performance stability during sensor disconnections. Nevertheless, both centralized neuro-evolution and decentralized neuro-evolution surpass the EFD strategy in smaller effectiveness degradations and reduced performance variability. This study provides insights into the performance of neuro-evolution under communication failures, especially for the respective robustness advantages of the centralized and decentralized control schemes, contributing to the development of more resilient urban drainage systems.