Reproduce “Two-Stage Active Distribution Network Voltage Control via LLM-RL Collaboration: A Hybrid Knowledge-Data-Driven Approach” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xu Yang, Chenhui Lin, Xiang Ma, Dong Liu, Ran Zheng, Haotian Liu, Wenchuan Wu
The authors propose a two-stage system for simulated grid voltage control: an LLM schedules OLTCs and capacitors from hourly regional forecasts, while an RL controller adjusts PV inverters using node-level measurements. The LLM policy improves with feedback, and RL pretraining precedes adaptation to its schedules. In 75 simulated IEEE 33-bus test episodes, mean voltage deviation was 8.22×10⁻³ p.u. and violation rate 2.60×10⁻³%.
Key insights
- LLMs can effectively process heterogeneous information and integrate semantic constraints directly into control policy formulation for complex power system dispatch tasks, moving beyond auxiliary roles.
- A Reflexion-based self-evolution mechanism for the LLM agent, combined with a similarity-based knowledge base, accelerates policy improvement and sample efficiency by providing richer feedback than scalar rewards.
- Pretraining the RL agent on diverse random settings and then finetuning with LLM-generated day-ahead schedules significantly enhances its adaptability and convergence for intra-day refinement tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.