Repository logo
Log In(current)
  1. Home
  2. Colleges & Schools
  3. Graduate School
  4. Doctoral Dissertations
  5. Trustworthy DRL-Based Real-Time Control for Urban Stormwater Systems: High-Fidelity Surrogates, Explainability, and Natural Flow Regime Restoration
Details

Trustworthy DRL-Based Real-Time Control for Urban Stormwater Systems: High-Fidelity Surrogates, Explainability, and Natural Flow Regime Restoration

Date Issued
May 2026
Author(s)
Fletcher, Isidora
University of Tennessee at Knoxville
Advisor(s)
Dr. Anahita Khojandi
Additional Advisor(s)
Dr. Jon Hathaway
Dr. Augustin Engman
Dr. Branko Kerkez
Permanent URI
https://trace.tennessee.edu/handle/20.500.14382/57587
Abstract
This dissertation explores the application of Deep Reinforcement Learning (DRL) for sustainable stormwater and pond management. Facing increasing strain on urban water systems due to rapid urbanization and climate change, Real-Time Control (RTC) emerges as a promising solution. However, its deployment is hindered by high computational costs, a lack of interpretability, and the ecological degradation of receiving streams. This research utilizes the U.S. EPA’s Storm Water Management Model (SWMM) to establish a comprehensive framework bridging the gap between theoretical DRL agents and practical civil engineering deployment. First, to address the computational bottleneck of high-fidelity hydraulic solvers, a Bidirectional Long Short-Term Memory (BiLSTM) surrogate model was developed (Chapter 2). This surrogate achieved a Nash-Sutcliffe Efficiency greater than 0.96 and enabled a 33-fold acceleration in DRL training. This efficiency allowed an Advantage Actor-Critic (A2C) agent to converge to a competent policy within a feasible timeframe, whereas direct simulation failed due to data starvation. Second, the “black box” problem of DRL was addressed using a novel SurrogateAccelerated Shapley Value Estimation (SVERL) framework (Chapter 3). This method moves beyond policy explanation to performance explanation, quantifying the “Value of Information” for sensor networks. The framework serves as a forensic tool, successfully diagnosing policy collapse and identifying structural conflicts between upstream storage and downstream erosion constraints. Finally, a novel Eco-Mimicry control framework was evaluated to address the Urban Stream Syndrome (Chapter 4). By developing a Natural Proxy model—a modified simulation representing pre-development hydrological conditions—this research established a dynamic ecological baseline. Experimental evaluation reveals that while the computational architecture is sound, the efficacy of eco-mimicry is fundamentally bounded by a Hydraulic Capacity Gap. The results demonstrate that intelligent control can optimize release timing but cannot overcome the physical storage deficits of gray infrastructure during peak flow events. Collectively, these studies provide a scalable, transparent, and ecologically conscious pathway for deploying autonomous control in sensitive urban watersheds.
Subjects

Reinforcement Learnin...

Stormwater Management...

Surrogate Models

Explainable AI

Adversarial Learning

SWMM

Disciplines
Data Science
Degree
Doctor of Philosophy
Major
Data Science and Engineering
File(s)
Thumbnail Image
Name

FletcherDiss_2026.pdf

Size

7.95 MB

Format

Adobe PDF

Checksum (MD5)

2277cfb109c0c4489b8f37be2c921509


University Libraries

1015 Volunteer Boulevard
Knoxville, TN 37996
865-974-4351

Map & Directions
Donate to the Libraries
  • About
  • John C. Hodges Society
  • Speaking Volumes magazine
  • Outreach
  • Directory
  • Employment
  • Policies
  • Library Intranet
University of Tennessee power T logo

The University of Tennessee, Knoxville
Knoxville, Tennessee 37996
865-974-1000

Events
A-Z
Apply
Privacy
Map
Directory
Give to UT
Accessibility

Built with DSpace-CRIS software - Extension maintained and optimized by 4Science