District cooling system control for providing regulation services based on safe reinforcement learning with barrier functions

• Published in: Applied energy Vol. 347; p. 121396
• Main Authors: Yu, Peipei, Zhang, Hongcai, Song, Yonghua
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2023
• Subjects: Control barrier function; control methods; District cooling system; dynamic models; energy; energy use and consumption; Frequency regulation service; Safe deep reinforcement learning; temperature; District cooling system; Safe deep reinforcement learning; Frequency regulation service; Control barrier function
• DOI: 10.1016/j.apenergy.2023.121396
• ISSN: 0306-2619

Thermostatically controlled loads (TCLs) in buildings are ideal resources to provide regulation services for power systems. As large-scale and centralized TCLs with high efficiency and large regulation capacity, district cooling systems (DCSs) have attracted great research attention for minimizing energy costs, but little on providing regulation services. However, controlling a DCS to provide high-quality regulation services is challenging due to its complex thermal dynamic model and uncertainties from regulation signals and cooling demands. To fill this research gap, we propose a novel safe deep reinforcement learning (DRL) control method for a DCS to provide regulation services. The objective is to adjust the DCS’s power consumption to follow real-time regulation signals subject to buildings’ temperature comfort constraints. The proposed method is model-free and adaptive to uncertainties from regulation signals and cooling demands. Furthermore, the barrier function is combined with traditional DRL to construct a safe DRL controller, which can not only avoid unsafe explorations during training (this may result in catastrophic control results) but also improve training efficiency. We conducted case studies based on a realistic DCS to evaluate the performance of the proposed control method compared to traditional methods, and the results demonstrate the increased effectiveness and superiority of the proposed control method. [Display omitted] •District cooling systems (DCSs) are controlled by a model-free reinforcement learning to provide regulation services.•A Gaussian process is adopted to learn the unknown system model to formulate safe sets.•A control barrier function is proposed to guarantee the constraint safety.•A neural-network-based method is proposed to improve computation efficiency.