TY - GEN
T1 - Transfer Learning with TD3 for Adaptive HVAC Control in Diverse Building Environments
AU - Kadamala, Kevlyn
AU - Chambers, Des
AU - Barrett, Enda
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
PY - 2025
Y1 - 2025
N2 - This paper studies the application of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm on two heterogeneous transfer scenarios. Transfer learning has shown to be effective in addressing challenges faced in RL for HVAC control by leveraging knowledge acquired during the development of an agent for one building to tackle a problem related to another building. However, buildings exhibit significant variability in size, construction materials, and geographical location; thus, simply transferring neural networks would be a challenge because of the need to adapt to diverse building characteristics. In this research, we extend prior work and investigate the efficacy of transfer learning with the TD3 algorithm. We use this algorithm to optimise HVAC control systems across different building environments. Our experimental results demonstrate the competitive performance of our transfer learning methods compared to rule-based control and training from scratch. Our transfer learning methods see up to 2–3% improvement in performance when compared to these agents. Overall, this study highlights the potential of transfer learning with the TD3 algorithm to enhance adaptive HVAC control systems in diverse building environments.
AB - This paper studies the application of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm on two heterogeneous transfer scenarios. Transfer learning has shown to be effective in addressing challenges faced in RL for HVAC control by leveraging knowledge acquired during the development of an agent for one building to tackle a problem related to another building. However, buildings exhibit significant variability in size, construction materials, and geographical location; thus, simply transferring neural networks would be a challenge because of the need to adapt to diverse building characteristics. In this research, we extend prior work and investigate the efficacy of transfer learning with the TD3 algorithm. We use this algorithm to optimise HVAC control systems across different building environments. Our experimental results demonstrate the competitive performance of our transfer learning methods compared to rule-based control and training from scratch. Our transfer learning methods see up to 2–3% improvement in performance when compared to these agents. Overall, this study highlights the potential of transfer learning with the TD3 algorithm to enhance adaptive HVAC control systems in diverse building environments.
KW - Continuous HVAC control
KW - Reinforcement learning
KW - Transfer learning
UR - https://www.scopus.com/pages/publications/85216094495
U2 - 10.1007/978-3-031-73058-0_21
DO - 10.1007/978-3-031-73058-0_21
M3 - Conference Publication
AN - SCOPUS:85216094495
SN - 9783031730573
T3 - Communications in Computer and Information Science
SP - 256
EP - 267
BT - Highlights in Practical Applications of Agents, Multi-Agent Systems, and Digital Twins
A2 - González-Briones, Alfonso
A2 - Julian Inglada, Vicente
A2 - El Bolock, Alia
A2 - Marco-Detchart, Cedric
A2 - Jordan, Jaume
A2 - Mason, Karl
A2 - Lopes, Fernando
A2 - Sharaf, Nada
PB - Springer Science and Business Media Deutschland GmbH
T2 - International Workshops on Practical Applications of Agents and Multi-Agent Systems, PAAMS 2024
Y2 - 26 June 2024 through 28 June 2024
ER -