Towards scalable implementation of advanced control strategies for HVAC systems in buildings

The building sector is one of the largest consumers of final energy, with HVAC systems representing a major share of this demand. Traditional rule-based HVAC controllers often prove inefficient under real operating conditions, as they tend to prioritize guaranteed indoor comfort over energy savings. To address this, recent research has focused on advanced control strategies such as MPC and DRL, which leverage real-time predictive capabilities to improve energy efficiency and flexibility. However, their adoption is limited by high computational requirements, often incompatible with existing BACS and by their "black-box" nature, which reduces stakeholder trust and hinders real-world implementation. This work is positioned within this context, aiming to facilitate the implementation of advanced controllers while improving their transparency and scalability. The HVAC system under investigation is a heating system consisting of an air–water source HP, a hot-water TES tank (which also acts as a hydraulic separator between the primary loop (HP-TES) and the secondary loop (TES–room distribution), and fan coil units as terminal devices in a single-zone building configuration. All control strategies operate without relying on indoor environmental state information, enhancing scalability to multi-zone applications, as the controller remains independent of the secondary-loop configuration. The proposed methodology begins by developing a DRL controller based on the SAC algorithm under a reference climate and occupancy schedule; an optimization pipeline is defined to identify the optimal configurations of the DRL in terms of learning hyperparameters and reward function weights. Among the optimized cases, a subset comprising those with the best performance is selected. From each of these cases, a decision tree model is developed to extract control rules from DRL policies and convert them into a rule set with an if–else format. The rule sets were ensembled and evaluated on the reference case. From this evaluation, the Ensemble-RE policy is extracted into a single if–else rule set (mirroring the procedure applied to the DRL controllers) thereby enhancing interpretability and transparency of the model. The resulting RE controller is subsequently tested on new scenarios with different climate and occupancy profiles to evaluate its performance and generalizability. To evaluate its performance, the RE controller is compared to a baseline RBC controller and to the best-performing DRL controller (developed from scratch and optimized for each scenario). In the reference case, the RE approach similarly to the DRL controller’s performance outperforming the baseline RBC in both energy consumption and daily peak power demand. Across various climatic and occupancy scenarios, RE methods consistently beat the baseline, though they are slightly less efficient than models trained specifically for these conditions. The result RE shows strong transferability, particularly in scenarios that deviate significantly from the reference. In all tested cases, it achieves at least a 6.9% reduction in energy consumption and at least a 22.7% reduction in daily peak power demand relative to baseline. These findings show that RE-based controllers can mimic advanced control policies while remaining simple, interpretable, and robust under varying boundary conditions.