Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Multi-source Transfer Learning Method for Enhancing the Deployment of Deep Reinforcement Learning in Multi-Zone Building HVAC Control
Deep reinforcement learning (DRL) control methods show great potential for optimal control of building HVAC systems, but require considerable time and data to learn effective policies. By employing transfer learning (TL) with pre-trained models, the need to learn the underlying data from scratch is avoided, thus saving time and resources. However, such a solution has the main critical issue: the inefficient utilization of multi-source domain control experience. Therefore, in this study, a multi-source transfer learning and deep reinforcement learning (MTL-DRL) integrated framework is proposed to utilize the control experience from different source domains for efficient building HVAC system control. The well-pretrained DRL parameters from the optimal multi-source transfer set are sequentially transferred to the target DRL controller. The results of a series of transfer experiments between buildings with different thermal zones and weather conditions indicate that the proposed MTL-DRL framework significantly reduces the training time of HVAC control, with improvements of up to 20% compared to DRL baseline models trained from scratch. Additionally, the MTL-DRL control method leads to a reduction in average energy consumption ranging from 1.43% to 3.12%, and a decrease in the average temperature deviation of up to 14.32%. Overall, this framework presents a promising solution for enhancing DRL-based methods in HVAC control performance by reducing training time and energy consumption while maintaining occupants’ comfort.
Multi-source Transfer Learning Method for Enhancing the Deployment of Deep Reinforcement Learning in Multi-Zone Building HVAC Control
Deep reinforcement learning (DRL) control methods show great potential for optimal control of building HVAC systems, but require considerable time and data to learn effective policies. By employing transfer learning (TL) with pre-trained models, the need to learn the underlying data from scratch is avoided, thus saving time and resources. However, such a solution has the main critical issue: the inefficient utilization of multi-source domain control experience. Therefore, in this study, a multi-source transfer learning and deep reinforcement learning (MTL-DRL) integrated framework is proposed to utilize the control experience from different source domains for efficient building HVAC system control. The well-pretrained DRL parameters from the optimal multi-source transfer set are sequentially transferred to the target DRL controller. The results of a series of transfer experiments between buildings with different thermal zones and weather conditions indicate that the proposed MTL-DRL framework significantly reduces the training time of HVAC control, with improvements of up to 20% compared to DRL baseline models trained from scratch. Additionally, the MTL-DRL control method leads to a reduction in average energy consumption ranging from 1.43% to 3.12%, and a decrease in the average temperature deviation of up to 14.32%. Overall, this framework presents a promising solution for enhancing DRL-based methods in HVAC control performance by reducing training time and energy consumption while maintaining occupants’ comfort.
Multi-source Transfer Learning Method for Enhancing the Deployment of Deep Reinforcement Learning in Multi-Zone Building HVAC Control
Lecture Notes in Civil Engineering
Francis, Adel (Herausgeber:in) / Miresco, Edmond (Herausgeber:in) / Melhado, Silvio (Herausgeber:in) / Hou, Fangli (Autor:in) / Ma, Jun (Autor:in) / Kwok, Helen H. L. (Autor:in) / Cheng, Jack C. P. (Autor:in)
International Conference on Computing in Civil and Building Engineering ; 2024 ; Montreal, QC, Canada
Advances in Information Technology in Civil and Building Engineering ; Kapitel: 7 ; 87-101
04.03.2025
15 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch