leochang017/microgrid-llm-coordination

A research simulator for studying whether a population of LLM agents (one per household) can negotiate peer-to-peer energy sharing during outages to be fair across heterogeneous households, robust to incomplete/incorrect info, and explainable to residents. The project is CS/ML-focused (coordination, robustness, explainability), not a power-systems optimizer. Who uses it: researchers and developers building LLM-based coordination methods and benchmarking them against heuristic and LP baselines. Key features: - Phase 1 simulator + Phase 1.6 hardening complete (CLI, 96 tests). - Pluggable coordination API: decide_transfers(t, …) -> list[Transfer]; Phase 2 will add LLM agents. - Household physics, neighborhood comm-graph, affiliation/overlay cliques, and settlement with caps/losses. - Data adapters: synthetic, Pecan Street, NREL, ResStock support; scripts to fetch data. - Baseline strategies: no_coordination, round_robin, round_robin_overlay, lp_optimal. - Scenario YAML configs, JSONL logging, metrics (served fraction, Gini, wasted/unmet), stress scenarios, and comparison tooling (gap_closed vs LP ceiling). Licensed MIT.