An intelligent model for managing visual comfort and building energy consumption through facade control using reinforcement learning

Abstract

Building facade characteristics have a significant influence on energy consumption for lighting and on occupant visual comfort, yet these two objectives frequently conflict with each other. Increasing the amount of incoming daylight reduces energy demand while simultaneously raising the risk of glare. Most prior studies have addressed these objectives in isolation or through offline optimization, and relatively few have focused on learning an online control policy that is sensitive to both goals simultaneously. In this study, a Deep Q-Network (DQN) agent was designed to control a three-state building facade (closed, half-open, and open). The reward function incorporated, in addition to the Energy Use Intensity (EUI) index, an explicit penalty term for the Daylight Glare Probability (DGP) index. The model was trained on real climatic data from three cities with notably different climates: Tehran, Mashhad, and Tabriz, in order to examine the generalizability of the learned policy across diverse solar conditions. Evaluation on the test dataset showed that the trained agent reduced hourly energy use intensity from the baseline of 12 W/m² to an average of 7.7223 W/m², predicted the optimal facade state with an overall accuracy of %96.465, and brought the mean DGP index to 0.0310, compared to 0.130 in the uncontrolled scenario. A comparison with prior studies indicated that the proposed method differs from similar work in its simultaneous coverage of multiple climates within a single model and its use of online decision-making rather than post-hoc optimization. This study presents a facade control framework whose energy and glare outcomes, at this stage, hold only within the simplified simulation setting described above; extending it toward real building deployment is left for future work.