Argonne demonstrates agentic AI to automate grid‑planning workflows

Argonne National Laboratory researchers demonstrated an "agentic" artificial intelligence system that can translate planning intent expressed in natural language into scenario setups, submit unit‑commitment solver jobs and produce written summaries, the presenters said.

Feng Chu, who identified himself as being "from Argonne National Laboratory," told attendees the project focuses on reducing the data‑cleanup and communication costs that make grid planning time‑consuming. "The agentic AI could be very useful here," Chu said, arguing the approach can automate repetitive workflows while leaving technical judgment to human planners.

The demonstration, led by a presenter identified in the session as Wigao, used Argonne's web‑based case builder for unit‑commitment studies. "Jl, which is Argonne's open source package, for security constraints, unit commitment, and production cost modeling," Wigao said during the demo, showing a three‑pane interface with a workspace, case editor and an agent panel that issues native API calls to the package rather than passing full raw JSON into the language model.

Presenters said that design reduces the risk of hallucination while preserving solution quality: by calling numerical solvers and exposing only high‑level statistics to the language model, the agent can rely on established solvers for numeric accuracy. Chu framed this as a tradeoff among approaches, noting that off‑the‑shelf large language models can hallucinate and that foundational models remain early research efforts; solver‑driven interfaces provide higher reliability but require coding interfaces. He recommended safeguards including code inspection, validation and sandboxing.

In the demo the agent performed complex, multi‑step tasks — ranking the most heavily loaded buses, applying randomized load adjustments, creating multiple scenario variants, and writing scripts that were submitted asynchronously to a solver. Wigao then opened result logs and dashboards that show iteration history, model size and final status, and used a map view to step through time and visualize congestion and locational marginal price differences.

When an audience member, Rick Benitez of BNGC, asked whether the tool handles seasonality, contingency analysis and production‑cost or capacity‑expansion modeling "—is that one of the things that the agents can do and look at?" —Chu replied that those functions are already implemented in the package. "These are already, like, implemented, like, in the package already. So it's a matter just activate veil. So your description in natural language will activate the agent to call those function, to perform those study," he said, and offered to check details after the session.

Presenters emphasized that the system is intended to assist, not replace, operators and planners by taking over tedious tasks and producing reviewable outputs (they said the agent writes results to a markdown file for human review). They also noted the regulatory context for planning work, referencing procedures and guidance from NERC and the Federal Energy Regulatory Commission (FERC), and argued that agentic AI is well‑suited for tasks that follow established protocols.

The demonstration highlighted potential efficiency gains for planners but also left implementation details to be confirmed: presenters said code, documentation and benchmark instances are publicly available online, and they invited follow‑up questions. The session concluded with a room‑wide Q&A and offers to follow up on specific modeling and cost‑input questions.