First-hand information for everyone
This paper argues that large language models (LLMs) and agentic AI should not be allowed to invent numerical answers for power systems. The authors show that LLMs can produce numbers that look plausible but break physical laws or operational limits. To avoid dangerous mistakes, they propose a simple rule: only report a numerical result when it comes from a trusted solver and passes an explicit verification check.
The authors present a tutorial-style guide to building such “solver-grounded” agentic systems for smart grids. They review practical building blocks — prompting strategies (for example, structured prompts, few-shot examples, chain-of-thought, and retrieval-augmented generation) and agent designs (including ReAct, plan-and-act, hierarchical and multi-agent patterns). They then test the idea in four case studies: wind power forecasting, electric vehicle (EV) charging scheduling, power flow analysis, and contingency diagnosis. Each case compares an LLM-only baseline to a solver-grounded system on the same data and metrics.
At a high level the proposed system divides labor into three roles. The LLM acts as an interpreter and orchestrator: it reads the request, gathers context, plans which tools to call, and writes explanations. Trusted solvers and simulators do the actual numerical work, like optimal power flow or forecasting. A verification gate checks the solver outputs against physical laws and operational limits before any numbers reach an operator. If verification fails, the agent either retries with bounded repairs or returns a safe failure message instead of fabricating values.
The experiments suggest concrete gains from this design. For EV charging, the EVAgent reproduced the CVXPY optimum and cut the unmet energy from the LLM-only approach by about 7.5–9.5 times. For contingency diagnosis, the GridDebugAgent repaired 17 of 39 problematic cases and reduced total violations by 52.3%. The paper also proposes a four-group evaluation framework to judge such systems on task utility, solver-grounded correctness, faithfulness and safe failure, and cost and latency.
Important caveats remain. The literature still lacks a consistent way to design and evaluate these systems, and verification itself is a hard research problem. LLMs have no built-in guarantee to respect physical constraints, so the system depends on the choice and reliability of external solvers and on the design of the verification gate. The authors highlight open questions including verification design, automated repair policies, standardized agentic benchmarks, and safer deployment practices. Their results are promising on the studied examples, but broader validation and careful engineering will be needed before such systems are used in live grid operations.