Mitigation of Flooding in Stormwater Systems Utilizing Imperfect Forecasting and Sensor Data with Deep Deterministic Policy Gradient Reinforcement Learning

• DOI: 10.20944/preprints202010.0413.v1
• Publication Date: 2020-10-21T06:37:21Z
• Authors: Sami Saliba; Benjamin Bowes; Stephen Adams; Peter Beling; Jonathan Goodall
• Keywords: Flood control; Implementation

Climate change and development have increased urban flooding, requiring modernization of stormwater infrastructure. Retrofitting standard passive systems with controllable valves/pumps is promising, but requires real-time control (RTC). One method automating RTC reinforcement learning (RL), a general technique for sequential optimization in uncertain environments. The notion that an RL algorithm can use inputs flood data rainfall forecasts to learn policy controlling the infrastructure minimize measures flooding. In real-world conditions, other state information, are subject noise uncertainty. To account these characteristics problem data, we implemented Deep Deterministic Policy Gradient (DDPG), distinguished by its capability handle input data. DDPG implementations were trained tested against policy. Three primary cases studied: (i) perfect (ii) imperfect forecasts, (iii) water level forecast Rainfall episodes (100) caused flooding system selected from 10 years observations Norfolk, Virginia, USA; 85 randomly used training remaining 15 unseen served as test cases. Compared system, all reduced volume 70.5% on average, performed within range 5%. This suggests robust noisy which essential knowledge advance applicability RTC.