LLMs and Agentic AI Systems for Smart Grids: A Tutorial on Architectures and Applications

Large Language Models (LLMs) and agentic AI systems are increasingly applied to power system tasks like forecasting, optimization, and contingency analysis. However, a significant gap exists between the fluent, human-like reasoning of LLMs and the rigorous requirements of grid operations. LLMs are probabilistic models that can generate numerically plausible but physically impossible outputs, such as dispatch schedules that violate Kirchhoff’s laws or thermal limits. Because these errors are often indistinguishable from correct results without manual inspection, they pose severe risks to grid safety and reliability.

This paper introduces a "solver-grounded" design principle to bridge this gap. Instead of treating the LLM as a source of truth, the architecture separates the system into three distinct roles:

1. The LLM (Orchestrator): Interprets natural language requests, plans multi-step workflows, retrieves relevant data, and summarizes results.
2. The Trusted Tool (Computer): Performs actual numerical computations using established power systems software (e.g., CVXPY, MATPOWER).
3. The Verification Gate (Guardrail): Explicitly checks the tool's output against physical and operational constraints before any information is presented to the operator.

This approach ensures that the LLM remains in the loop for its reasoning capabilities while delegating safety-critical calculations to verified, deterministic solvers.

The authors demonstrate this principle through four case studies: wind power forecasting, EV charging scheduling, power flow analysis, and contingency diagnosis. In each case, the solver-grounded agent significantly outperformed an LLM-only baseline. For instance, the "EVAgent" reproduced optimal results from a standard solver while reducing unmet energy by 7.5–9.5x compared to an LLM-only approach. Similarly, the "GridDebugAgent" successfully repaired 17 out of 39 contingency cases, reducing total grid violations by over 50%.