DARC: A Constraint-Diagnostic LLM Agent Framework for Day-Ahead Dispatch of Campus-Level Integrated Energy Microgrids Under Natural-Language Preferences and Forecast Uncertainty

Abstract

Day-ahead dispatch of integrated energy systems (IESs) is commonly solved by mixed-integer linear programming when objectives, constraints, and forecasts are fully specified. In practice, however, operators often express temporary preferences in natural language, and day-ahead forecasts inevitably deviate from realized operation. These two conditions make a fixed optimization interface difficult to use without additional modeling effort. This paper proposes DARC (Decompose-Act-Repair-Critique), a constraint-diagnostic LLM agent framework for day-ahead dispatch of campus-level microgrid-type IES with under-specified operating requirements. DARC combines three language modules, namely a Decomposer for temporal structure, a Resolver for numerical schedule generation, and a Critic for root-cause diagnosis, with deterministic repair, constraint checking, and metric evaluation. In all main experiments, the Decomposer is instantiated by its deterministic rule-based variant for reproducibility, so the reported results reflect a loop with two LLM modules (Resolver and Critic). The Projector repairs candidate schedules where possible, while the Checker supplies mathematical facts that ground the Critic’s feedback. DARC does not model forecast uncertainty through sets or scenarios; robustness to forecast deviation is pursued operationally, through margin-aware repair and checker-grounded revision, and is assessed empirically. Experiments on a campus-level microgrid IES testbed show that DARC can incorporate natural-language preferences, improve realized feasibility under forecast noise relative to both a point-forecast MILP and an interval-robust MILP baseline in terms of strictly feasible configurations, at higher operating cost, and produce diagnostic feedback that is robustly more useful than an ungrounded LLM diagnoser under judge models from three families, and at least as useful as checker facts alone, with its advantage concentrated in actionable revision guidance. DARC is therefore positioned not as a replacement for MILP, robust formulations, or model predictive control on fully specified problems, but as a complementary interface for operational settings where preferences and forecasts are not completely formalized.