著者

工藤 和樹 森 太郎

出版者

日本建築学会

雑誌

日本建築学会環境系論文集 (ISSN:13480685)

巻号頁・発行日

vol.84, no.759, pp.515-524, 2019 (Released:2019-05-30)

参考文献数

16

被引用文献数

2

---

Introduction The number of residences with thick insulation and air-tightness continues to increase in cold region. Ventilation loads cause over half of the heating load of such houses. In the previous report1), we developed a ventilation preheating system combining the solar thermal collector and the PCM panel shown in Fig. 1. We verified its performance with the experiment and numerical simulation. As the result, it is difficult to optimize the control system by a simple feedback system against ever-changing weather and indoor and outdoor environments. Therefore, in this research, we aim to develop a control method for a fan air volume of ventilation preheating system by machine learning. Initially, we outlined machine learning and reinforcement learning. Next, we explained how to introduce reinforcement learning in existing systems. Finally, we examined the performance of existing ventilation preheating system (VP system) and ventilation preheating system (RL system) controlled by reinforcement learning and examined the possibility of practical application.  Introduction of Reinforcement learning Fig. 2 and Fig. 3 shows the outline of reinforcement learning. We used the Q-Learning algorithm as the method of reinforcement learning. Fig. 4 shows the calculation process and simplified code of Q-Learning. The calculation formulas and algorithms were used to install reinforcement learning into the existing system. Fig. 6 shows a flowchart of the Q-Learning algorithm used in the calculation process in this study.  Control of ventilation preheating system with reinforcement learning We set the target schedule from October to March and prepared several cases of the operating periods and the air volume in VP system. We compared those results with the results operated by RL system. The findings obtained by this study are shown below. (1) As shown in Fig. 7 and 8, it was confirmed that the RL system automatically controlled the air flow of the fan. (2) As shown in Fig. 9, the RL system is equal to or higher than the performance of the VP system through the calculation period. Especially in the winter season, December to February, the performance of the RL system was the best in all cases. (3) Since reinforcement learning was successfully introduced. Also, the performance of RL system was better than the other cases, it is possible to study for practical application in the future.