Skip to content

Agentic AI Systems in Electrical Power Systems Engineering: Current State-of-the-Art and Challenges

🕒 Published (v1): 2025-11-18 13:18 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This review paper establishes a formal taxonomy distinguishing agentic AI from classical AI agents and generative AI, then presents four state-of-the-art case studies deploying agentic frameworks in electrical power systems engineering. It also catalogs failure modes and mitigation strategies for deploying trustworthy autonomous agents in safety-critical engineering domains.

Problem

No comprehensive study existed on the specific applications, challenges, and opportunities of agentic AI within electrical engineering. Prior surveys cover broad cross-industry use cases but lack depth on domain-specific deployment concerns (verification, tool injection, cascading misinformation, HITL governance) in safety-critical power systems workflows.

Method

The paper proceeds in three layers: (1) taxonomic formalization—agentic AI is characterized along seven dimensions (initiation, autonomy, task scope, learning, coordination, memory, architecture) distinguishing it from agent-based AI and generative AI; (2) four engineering case studies using orchestrator–delegator–specialist-agent architectures, two of which leverage MCP (Model Context Protocol) for tool integration; and (3) failure-mode analysis covering adversarial misinformation spread, agentic collusion, tool injection, and model ambiguity, paired with mitigation recommendations including zero-trust architecture and HITL governance.

Key Contributions

  • First comprehensive review specifically connecting agentic AI theory to electrical/power systems engineering practice.
  • Formal taxonomy table mapping AI agents, generative AI, and agentic AI across seven architectural dimensions (initiation, autonomy, task scope, learning, coordination strategy, memory, core components).
  • Four detailed case studies: (a) MCP-based framework for power system studies and benchmarking, (b) MCP-based substation illumination study agent, (c) engineering bill-of-quantity (BoQ) generation from RFQ documents, (d) survival analysis agent for battery swapping station dynamic pricing.
  • Structured failure-mode investigation: adversarial false-information spread, cascading LLM misinformation, agentic collusion, tool/prompt injection, model ambiguity.
  • Actionable deployment recommendations including zero-trust architecture and HITL governance for safety-critical agentic systems.

Results

The paper is a review/survey; it reports qualitative outcomes from case studies rather than aggregate benchmark numbers. Specific quantitative results from the four case studies are not included in the provided text excerpt. The survey of prior work (Table 2) catalogs 9 third-party systems across automotive, 6G, urban planning, energy markets, and control tuning domains—none with directly comparable metrics in the excerpt.

Limitations

  • Quantitative benchmarking data for the four internal case studies is not presented in the provided text.
  • Case studies are evaluated in isolated, often laboratory-style settings; real-world grid deployment is not demonstrated.
  • The taxonomy is descriptive rather than formal (no mathematical grounding), making rigorous benchmarking across paradigms difficult.
  • Security mitigations (zero-trust, HITL) are discussed conceptually; empirical adversarial robustness evaluations are absent.
  • Coverage of multi-agent coordination protocols and formal verification is acknowledged as an open gap but not addressed.

Relevance to Harnesses / Meta-Harnesses

The orchestrator–delegator–specialist-agent architecture described here is functionally a meta-harness pattern: a top-level controller decomposes goals, routes sub-tasks to specialized agents, and consolidates results—exactly the structure implemented in software harnesses like LangChain, AutoGen, and MetaGPT (cited in the survey). The MCP-based case studies make this concrete: MCP acts as the tool-invocation layer that a harness wraps around domain simulators, APIs, and knowledge bases. The failure-mode analysis (tool injection, cascading misinformation, agentic collusion) directly informs the safety and governance requirements any meta-harness designer must build in, particularly for high-stakes or automated pipelines.