Trustworthy large language models (LLMs) for power-system operations: From "Grid Copilots" to verifiable decision support

Abstract

Large language models (LLMs) are increasingly proposed as "Grid copilots" for alarm interpretation, procedural retrieval, incident summarization, scenario analysis, and tool-mediated power-system studies. This review combines a PRISMA 2020-informed systematic identification/screening component with structured qualitative synthesis across power systems, LLM agents, trustworthy AI, and safety-critical decision support. The documented review pathway identified 312 records, removed 48 duplicates, screened 264 records, and assessed 213 full-text articles; after incorporation of additional recent and methodological literature, the current manuscript draws on a total evidence base of 199 references. The synthesis shows that recent tool-using agents can execute power-flow/OPF and contingency workflows, but operational trust remains conditional on authenticated state/topology, valid models and constraints, solver/tool correctness, uncertainty calibration, deadline compliance, human authority, cybersecurity, and reconstructable audit evidence. We therefore present VDS-LLM-Grid as an evidence-informed conceptual reference architecture-not a validated deployment-that integrates six assurance layers: input/state grounding, constraint-aware reasoning, independent verification, calibrated uncertainty with abstention, human oversight, and governance/auditability within a crosscutting cybersecurity boundary. The paper defines conditional "verifiable decision support," mathematical systemlevel metrics, deadline-aware routing, and a study-level evidence map that prioritizes empirical end-to-end validation, constraint verification, calibration, latency, cybersecurity, robustness, and operator evaluation.