Enhancing Stormwater Management through Machine Learning-based Real-time Prediction and Control

Many cities face high levels of flooding and pollution from stormwater runoff due to factors such as ongoing urbanization and aging stormwater management infrastructure. As climate change continues to alter precipitation, temperature, and sea levels, existing stormwater systems will be pushed beyond their designed capacity, further increasing flooding and pollution. This dissertation focuses on real-time prediction and control of stormwater related systems as a means to enhance community resilience to these issues. The research advanced the application of emerging deep machine learning techniques to water resources engineering using the coastal city of Norfolk, Virginia as a test-bed for these novel approaches. Norfolk faces recurrent flooding from storm events and ongoing sea level rise, while having to reduce polluted stormwater runoff entering the Chesapeake Bay. The first study uses supervised deep machine learning to create forecasts of groundwater table response to storm events, providing additional information for flood forecasting and stormwater management. The second and third studies explore deep reinforcement learning as a method for real-time control of stormwater systems. In the second study, reinforcement learning is used to create control strategies that mitigate flooding in a simple stormwater system scenario inspired by a watershed in Norfolk. The third study uses reinforcement learning for real-time stormwater system control with the competing objectives of mitigating flooding while also improving water quality by capturing sediment. This was done using a real-world simulation of Norfolk's Hague neighborhood instead of the simplified system from the second study. Key findings from this research are (i) deep machine learning can be used to create real-time hourly forecasts of the groundwater table response to storm events in a coastal city using forecast rainfall and tide conditions as input data with a mean root mean squared error of 0.05 m, (ii) reinforcement learning can learn real-time stormwater system control strategies that reduce flooding compared to conventional, uncontrolled stormwater systems by 32%, (iii) system-level stormwater real-time control with reinforcement learning can reduce flooding by 13% compared to local-scale control rules, and (iv) reinforcement learning can use real-time water quality observations to reduce sediment loads by an average of 52% with only a small increase in flooding (5%) compared to conventional, uncontrolled stormwater systems. While this dissertation has focused on coastal cities, the knowledge and methods developed could be applied to inland stormwater systems as well. These advancements contribute to a growing body of knowledge related to smart stormwater systems, which can aid communities through improved prediction and control of stormwater to reduce flooding and pollution.