AutoB2G: A Large Language Model-Driven Agentic Framework For Automated Building-Grid Co-Simulation

A research team from UNSW Sydney and CSIRO has developed AutoB2G, a novel framework that uses large language models (LLMs) to automate the complex process of building-grid co-simulation. The system addresses a critical gap in energy management: while reinforcement learning can optimize building operations, existing tools lack systematic evaluation of grid-level impacts and require substantial manual configuration. AutoB2G extends the CityLearn V2 environment to support building-to-grid interactions and employs the SOCIA (Simulation Orchestration for Computational Intelligence with Agents) framework to transform natural language task descriptions into complete simulation workflows.

The framework's innovation lies in its structured approach to guiding LLMs through complex simulation tasks. Since LLMs lack prior knowledge of simulation implementation contexts, the researchers constructed a comprehensive codebase organized as a directed acyclic graph (DAG) that explicitly represents module dependencies and execution order. This structure enables the LLM to retrieve complete executable paths, automatically generating, executing, and iteratively refining simulators. Experimental results demonstrate that AutoB2G can effectively coordinate building-to-grid interactions to improve grid-side performance metrics while dramatically reducing the programming expertise traditionally required for such simulations.