EUVIMEDEUVIMED entdecken
Crossref

Intrinsic Reward-Driven SAC-IRCNet: A Novel Energy-Saving Control Method for HVAC Systems

Jian Cen, Linzhe Zeng, Xi Liu, Xinyao Li, Yongkang Li, Jianming Yang, Chengming Huang, Feiqi Deng

IEEE Transactions on Automation Science and Engineering · 2026 · 23 · 9745-9757

DOI: 10.1109/tase.2026.3679538 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Automation Science and Engineering, 2026. Verfasst von Jian Cen und 7 weiteren Autor:innen. Die Arbeit nennt 33 Literaturangaben. Gefördert von: Natural Science Foundation of Guangdong Province, Natural Science Foundation of Guangdong Province of Offshore Wind Power Joint Foundation, National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the rising global energy consumption, the energy use of heating, ventilation, and air conditioning (HVAC) systems has become a critical concern. Existing deep reinforcement learning control methods for HVAC systems often exhibit slow convergence and poor adaptability to dynamic environments, resulting in significant indoor temperature fluctuations, inefficient temperature control strategies, and consequently, energy waste and failure to meet thermal comfort requirements. To address these challenges, this study proposes a novel HVAC control method based on the Soft Actor-Critic with Intrinsic Reward and Correlation-Aware CNN (SAC-IRCNet) model. The model incorporates cooling load prediction as a constraint to enable on-demand cooling supply. It integrates an Elliptical Dynamics Exploration intrinsic reward mechanism, which accelerates SAC convergence through elliptical exploration rewards and inverse dynamics models, thereby reducing energy waste. Additionally, the Correlation-Aware CNN enhances SAC’s feature extraction capability by leveraging state correlations to better understand contextual information, enabling more accurate responses to environmental changes and improved thermal comfort. Experimental results show that SAC-IRCNet achieves a 3.94% faster reward convergence and a 5.78% higher maximum cumulative reward within five episodes compared to SAC. It reduces energy consumption by up to 17.25% (21,837.58 kW) and lowers thermal discomfort violations by up to 6.44%, demonstrating excellent generalization ability across two datasets. Note to Practitioners—This study proposes a new control method based on the SAC-IRCNet model to address the energy efficiency and thermal comfort issues of heating, ventilation, and air conditioning (HVAC) systems. The conventional deep reinforcement learning methods used for HVAC control, such as DQN, PPO, DDPG, and TD3, usually have some problems. First, value function-based methods like DQN cannot directly handle continuous action spaces. Second, mainstream on-policy algorithms such as PPO suffer from low sample efficiency, leading to excessively high training costs in computationally expensive building simulations. More critically, widely used off-policy algorithms for continuous control—DDPG and its improved version TD3—exhibit limited exploration capabilities due to their deterministic policies, hypersensitivity to hyperparameters, and unstable training. These limitations often result in slow convergence, suboptimal policies, and potentially unstable control commands in HVAC applications, which can induce indoor temperature fluctuations. The proposed SAC-IRCNet model enhances the feature extraction ability through a convolutional neural network with correlation perception, integrates the cooling load prediction constraint to achieve on-demand cooling, and introduces an elliptical intrinsic reward mechanism to accelerate convergence. This method not only reduces energy consumption by up to 17.25%, but also improves thermal comfort by minimizing discomfort violations. Due to its excellent generalization ability, this method can be applied to various HVAC systems to achieve more efficient energy utilization.

Förderung

  • Natural Science Foundation of Guangdong Province
  • Natural Science Foundation of Guangdong Province of Offshore Wind Power Joint Foundation
  • National Natural Science Foundation of China

Zitieren

APA

Cen, J., Zeng, L., Liu, X., Li, X., Li, Y., Yang, J., Huang, C., & Deng, F. (2026). Intrinsic Reward-Driven SAC-IRCNet: A Novel Energy-Saving Control Method for HVAC Systems. IEEE Transactions on Automation Science and Engineering, 23, 9745-9757. https://doi.org/10.1109/tase.2026.3679538

Vancouver

Cen J, Zeng L, Liu X, Li X, Li Y, Yang J, et al. Intrinsic Reward-Driven SAC-IRCNet: A Novel Energy-Saving Control Method for HVAC Systems. IEEE Transactions on Automation Science and Engineering. 2026;23:9745-9757. doi:10.1109/tase.2026.3679538

BibTeX
@article{Cen2026,
  author = {Cen, Jian and Zeng, Linzhe and Liu, Xi and Li, Xinyao and Li, Yongkang and Yang, Jianming and Huang, Chengming and Deng, Feiqi},
  title = {Intrinsic Reward-Driven SAC-IRCNet: A Novel Energy-Saving Control Method for HVAC Systems},
  journal = {IEEE Transactions on Automation Science and Engineering},
  year = {2026},
  volume = {23},
  pages = {9745--9757},
  doi = {10.1109/tase.2026.3679538}
}
Kostenlos starten →