Designing Robust Control Rules for Stochastic Engineered Systems 107 views

Limited availability of in-situ hydrological data stands as a challenge to modeling and managing water resources systems. This is because effective water infrastructure planning and management requires an understanding of natural system processes and their interactions with the built environment; yet limited in-situ data requires the estimation of model parameters describing these processes through calibration. Such parametric model uncertainty can have significant implications for infrastructure design; however, it is often ignored in the design stage. In response to this pressing issue, this research proposes to improve the quantification and management of uncertainty in hydrological models. First, we introduce innovative diagnostic tools to assess the performance of Markov Chain Monte Carlo (MCMC) algorithms in calibrating complex physical models with high-dimensionality and multimodality using analytical test problems as benchmark examples. Second, we propose to utilize our knowledge of effective algorithms gained through the first study to quantify parametric uncertainty in a Stormwater Management Model (SWMM) of an urbanizing system. We then propose to design stochastic multi-objective control rules for flood risk reduction that are robust to this uncertainty. The latter step will contribute to the literature through two papers: The first study will introduce Evolutionary Multi-Objective Direct Policy Search (EMODPS) to the stormwater control literature and compare it with Deep Deterministic Policy Gradient (DDPG), which has been used in designing stormwater control rules. The second study will provide insights into how to most effectively design stormwater control rules that account for parametric hydrological model uncertainty. This research promises to advance our understanding of how to better quantify and manage uncertainty in water resources systems.

PHD (Doctor of Philosophy)

Bayesian Inference; MCMC; Multi Objective Reinforcement Learning; Robust Optimization

English

All rights reserved (no additional license for public reuse)

2024-07-24