有效的暖通空调操作不仅实现了节能，还为居住者创造了更舒适的室内环境。目前，商业建筑，尤其是开放式办公建筑面临着不必要的能源浪费和室内热环境共享不理想等问题。因此，开发一种新的暖通空调系统框架模式，使每个人都能在自己喜欢的热环境下工作，并实现更高的能源效率，具有重要意义。此外，即使乘客的行为和偏好对HVAC性能有相当大的影响，大多数针对乘客相关因素的传感系统可能会导致隐私问题或过于侵入。因此，有必要开发新的非侵入性、非私密的传感框架，用于实时监测个人热舒适性- 时间最后，提出了一种基于人机交互的空调制冷系统，以提高空调系统的舒适性。因此，本文提出并评估了两种交互式控制算法，并在常规共享办公空间中进行了现场对比研究。所提出的控制算法分别基于标准优化和强化学习。现场试验结果表明，采用标准优化控制器的控制算法可以最大限度地降低能耗，同时提高热舒适性，而基于强化学习的控制器仅依赖于乘员的反馈，难以覆盖所有可能的有限状态。 引用:2019年年度会议，密苏里州堪萨斯城，会议论文

The effective HVAC operations not only achieve energy savings but also create a more comfortable indoor environment for occupants. Currently, commercialbuildings, particularly open-plan office buildings are faced with problems like unnecessary energy waste and unsatisfactory shared indoor thermal environment.Therefore, it is significant to develop a new paradigm of HVAC system framework so that everyone could work under their preferred thermal environmentand the system can achieve higher energy efficiency. Moreover, even if occupant behaviors and preferences have considerable impacts on HVAC performances,most of sensing systems for occupant-related factors may either result in privacy issue or are too intrusive. Hence, it is necessary to develop new non-intrusiveand less private sensing framework for monitoring individual thermal comfort in real-time. Lastly, taken human-in-the-loop into consideration, the proposedsystem aims to optimize energy performance and improve occupant thermal comfort by developing interactive control logics for an integrative HVAC systemfeaturing personalized cooling with non-intrusive sensing techniques. Therefore, this paper proposes and evaluates two interactive control algorithms byconducting an in-situ and comparative study in a regular shared office space during cooling season. The proposed control algorithms are based on standardoptimization and reinforcement learning, respectively. The in-situ results have shown that control algorithms with standard optimization controller canminimize energy consumption while improve thermal comfort while the reinforcement learning-based controller has difficulty in covering all possible finitestates if the controller is only dependent on occupants’ feedbacks.