Skip to main content
Back to timeline
arXivSource publication:

A three-agent team now sits between buildings and the distribution grid, lifting task success from 38.0% to 69.0% on 200 synthesized tasks

Synopsis

Dispatch interactions between buildings and the distribution grid have long depended on specialists from both domains translating requests and hand-building simulation and optimization workflows. B2G-Agent splits that process across three collaborating agents: a Translator for bidirectional semantic alignment, an Actor that decomposes tasks into directed acyclic graphs and calls EnergyPlus and pandapower, and a Reviewer that combines semantic and evidence checks and routes failures back for correction. Across 200 synthesized interaction tasks, its interaction quality score rises from the baseline's 66.71 to 85.57 and task success from 38.0% to 69.0%, while operational constraints are satisfied.

Source-provided article image: B2G-Agent: An LLM-Based Multi-Agent Tool for End-to-End Building-Grid Interactions
arXiv

Interpretation

Dispatch interactions between buildings and the distribution grid can now be carried end to end by a set of large language model agents, with the distribution network operator and building energy managers supplying only task requirements and feedback. Previously, either each building energy manager submitted flexibility responses directly to the operator, or an aggregator coordinated buildings and reported their combined response, with problem formulation, simulation, optimization and verification done by hand by specialists from both fields. Across 200 synthesized interaction tasks, B2G-Agent reaches an interaction quality score of 85.57 against 66.71 for the single-agent baseline, and a task success rate of 69.0% against 38.0%.

Three agents divide the stages of a cross-domain interaction, keeping cross-field modeling and computation away from the users. The Translator uses role-conditioned hybrid retrieval-augmented generation to map one side's dispatch intent into instructions the other side can execute; the Actor applies backward task decomposition to trace target fields into a directed acyclic graph of tool calls and invokes EnergyPlus and pandapower for joint simulation or optimization; the Reviewer locates failures with semantic checks plus deterministic evidence validation and returns the task to the responsible agent. The study uses the IEEE 33-Bus Test System as the distribution network with 10 building energy managers, each represented by 1 to 5 EnergyPlus building prototypes, evaluates interactions over a day at 15-minute resolution, and covers seven task categories: peak shaving, congestion management, voltage management, reserve service, renewable integration, flexibility assessment and market trading.

Perspective

The evaluation covers 200 synthesized interaction tasks, with the IEEE 33-Bus Test System as the distribution network and 10 building energy managers each represented by 1 to 5 EnergyPlus building prototypes, assessed over a day at 15-minute resolution; the tool and interface are released as open source.

The evaluation tasks are synthesized, online memory is disabled during benchmark evaluation so the Reviewer retrieves only from the same read-only seed repository, and future work is planned to test effectiveness under practical validation on the basis of simulated tasks and to develop explanation functions for decisions made for both operators and building energy managers.