Graph-Based Adaptive Policy Learning for Multi-Zone HVAC Control

Abstract

Multi-zone heating, ventilation, and airconditioning (HVAC) control is important for sustainable smart buildings because it must balance thermal comfort and energy-aware operation under changing weather and occupancy conditions. This study presents a graph-based adaptive policy learning framework for synthetic multi-zone HVAC control and evaluates whether graph representation provides measurable benefits over a non-graph policy. A synthetic three-zone building is represented as a graph in which nodes denote thermal zones and edges encode predefined neighboring-zone relations. A two-layer graph convolutional network (GCN) policy is compared with a Flat multilayer perceptron (MLP) baseline using identical zone-level information, behavioral-cloning initialization, derivative-free parameter adaptation, and independent held-out evaluation. Results show that the GCN policy achieved the best comfort recovery under hotter tropical and high occupancy conditions, reducing cumulative comfort violation to $5.46 \pm 3.51$ and 3.17 ± 3.02, respectively. In contrast, the Flat MLP achieved the best performance under the milder-day scenario, with a cumulative reward of $\mathbf{- 2 1. 5 3} \pm \mathbf{0. 5 8}$ and comfort violation of $\mathbf{0. 2 6} \pm \mathbf{0. 1 3}$. These findings indicate that graph-based representation is not universally superior in a minimal three-zone setting, but can provide advantages under stronger inter-zone disturbances. The reported energy values are internal energy proxies rather than measured electricity consumption in kWh. This study provides controlled proof-of-concept evidence for adaptive graph-based HVAC control.