Statistical Learning Approaches For The Control Of Stormwater Systems

Abstract

Rapid advances in wireless communication, embedded systems, and high-performance computing are promising the fusion of physical and digital water. The next generation of stormwater systems --- equipped with wireless sensors and actuators --- will autonomously reconfigure themselves to prevent flooding and achieve system scale objectives. This vision of "smart'" stormwater systems is not limited by technology, which has matured to the point where it can be ubiquitously deployed. Instead, the challenge is much more fundamental and rooted in a system-level understanding of stormwater networks: once stormwater systems become highly instrumented, how should they be controlled to achieve the desired watershed outcomes? This dissertation leverages statistical learning methods to begin closing fundamental knowledge gaps in the emerging field of smart water systems. The second chapter of this dissertation addresses the lack of simulation tools for modeling pollutant interactions by introducing a new toolchain for coupling the existing hydraulic, hydrologic, and water quality models. Using this toolchain, we demonstrate real-time control's potential for enhancing nutrient removal in a watershed. In the third chapter, to characterize a watershed's controllability, a real-world case study is carried out using a wireless sensor-actuator network. Through this study, the ability to precisely shape the hydrograph is quantified, illustrating the high level of granularity that can be achieved using real-time control. Given that most state-of-the-art stormwater control algorithms require surrogate models or assume simplified dynamics, the fourth chapter introduces a Reinforcement Learning-based model-free algorithm for synthesizing stormwater controllers. The efficacy of the algorithm is demonstrated via simulation, highlighting strong performance. More importantly, a discussion is provided on the limitations of the approach, and a set of guidelines is presented for those seeking to apply Reinforcement Learning to stormwater control. The fifth chapter in this dissertation introduces a Bayesian Optimization-based methodology for addressing the lack of knowledge relating to the uncertainty in stormwater control approaches and demonstrates its use for identifying robust control strategies. In the final chapter, an open-source Python library to facilitate the systematic quantitative evaluation of control algorithms is introduced. This library provides a curated collection of stormwater control scenarios, coupled with an accessible programming interface and a stormwater simulator, to provide a standalone package for developing stormwater control algorithms. The discoveries made in this dissertation, along with the algorithms and tools developed, seek to support the development of a new generation of autonomous stormwater infrastructure.