Deployment Readiness Assessment of Large Language Models in Power Distribution Systems: A Risk-Aware Framework

Kul S., ARSLAN B., Celtek S. A.

8th Global Power, Energy and Communication Conference, GPECOM 2026, Naples, İtalya, 3 - 05 Haziran 2026, ss.1145-1150, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/gpecom70462.2026.11578628
  • Basıldığı Şehir: Naples
  • Basıldığı Ülke: İtalya
  • Sayfa Sayıları: ss.1145-1150
  • Anahtar Kelimeler: adversarial robustness, deployment readiness, hallucination, human-in-the-loop, Large language models, power distribution systems, risk assessment, smart grid
  • Gazi Üniversitesi Adresli: Evet

Özet

Large Language Models (LLMs) are increasingly being proposed for a wide range of power distribution system tasks, from fault diagnosis and state estimation to dispatch optimization and energy management coordination. However, deployment decisions in existing research are predominantly evaluated through task-performance metrics alone, overlooking the operational risks that arise when probabilistic generative systems interact with safety-critical physical infrastructure. This paper proposes the LLM Deployment Readiness Assessment (LDRA) framework, an analytical instrument that evaluates any LLM-integrated power system use case along four axes: Task Determinism Requirement (TDR), Hallucination Criticality Index (HCI), Adversarial Exposure Surface (AES), and Human-in-the-Loop Feasibility (HILF). A weighted Composite Readiness Score (CRS) maps each use case into one of three deployment zones (GREEN, AMBER, or RED), enabling structured go/nogo decisions independent of performance benchmarks. We apply LDRA to twelve representative use cases drawn from recent literature and demonstrate that four widely-studied deployments fall in the RED zone despite strong benchmark results or active research presence, revealing a systematic gap between reported performance and operational readiness. The framework provides power system engineers and AI practitioners with a shared vocabulary and a reusable analytical methodology for risk-aware LLM integration.