AI/ ai · energy · simulation · reinforcement-learning

AutoB2G Lets Researchers Describe Power Grid Tests in Plain English

A new agentic framework uses LLM agents to auto-build the complex co-simulations needed to study AI-controlled buildings on live power grids.

An AI framework called AutoB2G can turn a plain-English description into a working simulation of buildings, learning algorithms, and power grids — no manual wiring required.

Researchers at the paper's institution built AutoB2G to solve a specific bottleneck: studying how smart buildings respond to grid signals requires running at least three heterogeneous simulators in sync — one for the building, one for the reinforcement learning agent, and one for the distribution grid. Getting those simulators to talk to each other correctly has historically been a hand-coded slog. AutoB2G reframes that as a workflow orchestration problem: users state a high-level goal in natural language, and LLM agents handle retrieval, code composition, execution, verification, and iterative repair. The underlying codebase is organized as a directed acyclic graph, which keeps components modular and lets the system patch individual nodes when something breaks.

The practical upside is faster prototyping for a research area that actually matters. Demand-side flexibility — making buildings soak up or shed load on cue — is one of the cleaner levers for stabilizing grids increasingly stressed by renewable intermittency. A tool that lowers the barrier to simulating those dynamics could accelerate real policy and infrastructure decisions.

AutoB2G is an academic paper, not a shipping product, and LLM-generated simulation pipelines will need rigorous validation before anyone trusts them for grid planning — but as a scaffolding tool for researchers, the direction is credible.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →