AgentiControl-MPC: Automated model predictive control for buildings using large language model-based multi-agent system

Model predictive control (MPC) has demonstrated significant advantages in building energy system control. However, the development of MPC controllers remains expertise-intensive and time-consuming, requiring substantial domain knowledge in modelling, control formulation and software implementation, which limits large-scale deployment in practice. To address these challenges, this paper develops an agentic system, AgentiControl-MPC. It is a large language model (LLM)-based, multi-agent system that automates the development of MPC in buildings. AgentiControl-MPC consists of six specialized LLM-based agents that collaboratively perform tasks on data processing, predictive model design, MPC formulation, code generation, and error correction. Effective collaborations across agents are achieved by using well-designed communication mechanisms, which include inter-layer, intra-layer and state-aware communications. The effectiveness, reliability and robustness of AgentiControl-MPC in automating MPC controller development are demonstrated on the Building Energy Simulation Test (BESTEST) Hydronic Heat Pump test case from the Building Optimization Testing Framework (BOPTEST) through comparison with a single-agent LLM baseline. In repeated-run experiments, AgentiControl-MPC achieved a 93.3 % controller-generation success rate, compared with 66.7 % for the single-agent LLM baseline, and produced more consistent baseline-outperforming MPC controllers. The selected automatically generated MPC reduced energy cost by 52 % and thermal discomfort by 69 % compared with the baseline Proportional-Integral (PI) controller.