Virginia Tech
Reinforcement Learning Benchmarking for Sustainable Energy Systems: Perturbation Robustness, Safety Constraints, and Multi-Agent Coordination
Abstract
dc:description.abstractReinforcement learning (RL) has emerged as a promising approach for sequential decision making in sustainable energy systems, yet the lack of systematic benchmarks hinders principled algorithm comparison and deployment. This thesis presents SustainRL-Bench, a comprehensive benchmarking study that evaluates RL algorithms across three sustainable energy environments—electric vehicle (EV) charging coordination, building HVAC control, and cogeneration plant dispatch—along three experimental axes: perturbation robustness, safety constraints, and multi-agent coordination. Our evaluation spans 116 unique configurations totaling over 310 training runs. Several findings challenge conventional expectations. Perturbation sensitivity varies by over an order of magnitude across environments: EV Charging PPO degrades less than 11% under state perturbation, while Building policies collapse by over 100% at the same relative scale, and in Cogeneration, on-policy PPO remains far more robust than off-policy methods (<24% degradation across all channels vs. order-of-magnitude losses for SAC). Constrained MDP-based safe RL achieves precise constraint tracking only when the cost signal is sufficiently orthogonal to the reward; partial overlap permits tracking at tight limits, while full alignment renders constraints inactive, and off-policy Lagrangian methods fail entirely across all environments. In multi-agent settings, environment coupling structure rather than agent count determines algorithm feasibility, with the single-agent algorithm ranking reversing in tightly coupled domains. All code and configurations are publicly available.
Degree
thesis:*- Name thesis:degree_name
- Master of Science
- Level thesis:degree_level
- masters
- Discipline thesis:degree_discipline
- Electrical Engineering
- Department dc:contributor.department
- Electrical Engineering
- Grantor dc:publisher
- Virginia Tech
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Koruturk, Mehmet
- Chair dc:contributor.committeechair
-
- Jin, Ming
- Committee members dc:contributor.committeemember
-
- Soysal, Alkan
- Mehrizi-Sani, Ali
Subjects
dc:subject × 5Rights
dc:rights- Statement dc:rights
-
- In Copyright
- Licence dc:rights.uri
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Other
- vt_gsexam:46832