Causality-constrained hierarchical LLM agents for online rule adaptation in building HVAC control

Adaptive multi-zone HVAC control often requires identifying a building-specific predictive model or training a policy through environment interaction. These requirements increase the effort needed to deploy adaptive controllers across buildings with different HVAC configurations. To address this challenge, we propose a causality-constrained hierarchical large language model (LLM) agent framework for online cooling-setpoint rule adaptation. The framework operates in two stages. Offline onboarding standardizes building management system (BMS) points and establishes a human-confirmed causal graph. During online rule adaptation, hierarchical agents propose hourly updates to rule-based controllers (RBCs), while the RBCs generate 15-min setpoints. Structural and causal checks, together with zone cooling allowances, determine whether each proposed update is incorporated into the RBC. We evaluate the framework in three BOPTEST cases with one, two, and five zones against basic and enhanced RBCs, hierarchical model predictive control, proximal policy optimization (PPO), and multi-agent PPO. The framework achieved the highest mean reward and the lowest mean energy use among the compared controllers in all three cases. Energy use was 2.2–43.2% lower than that of the baselines, accompanied by small accumulated PMV deviations of 0.0192–0.4542 PMV ⋅ h. The causal experiments showed that the complete causal layer reduced accumulated PMV deviation and safety interventions in every case, primarily by constraining and stabilizing rule adaptation. Further ablations quantified the effects of memory, cross-zone coordination, and LLM thinking on rule adaptation, control performance, and LLM inference overhead. Overall, the framework combines competitive energy and comfort performance with explicit, reviewable online rule adaptation, supporting its potential for adaptive multi-zone HVAC control across the evaluated HVAC configurations.