A novel defrosting initiation strategy based on convolutional neural network for air-source heat pump. (August 2021)

• Record Type: Journal Article
• Authors: Wang, Wenyi; Zhou, Qun; Tian, Guanyu; Wang, Yikai; Zhao, Zhongfan; Cao, Feng
• Is Part Of: International journal of refrigeration. Volume 128(2021)
• Page Start: 95
• Page End: 103
• Publication Date: 2021-08
• Subjects: Air-source heat pump -- Defrosting initiation strategy -- Convolutional neural network -- Defrosting model -- Air source heat pump experiment; Pompe à chaleur aérothermique -- Stratégie de déclenchement du dégivrage -- Réseau neuronal convolutif -- Modèle de dégivrage -- Expérience sur une pompe à chaleur aérothermique; Refrigeration and refrigerating machinery -- Periodicals; 621.56
• DOI: 10.1016/j.ijrefrig.2021.04.001
• Languages: English
• ISSNs: 0140-7007

Abstract

Abstract: Defrosting is an essential and vital part of the air source heat pump (ASHP) to restore the heating capacity and operating performance during the wintertime when the humid ambient air can cause frosting on the surface of the evaporator. A timely and effective defrosting control strategy that determines the defrosting start and exit time is critical for improving the system operating performance. However, current time-based defrosting methods that make defrosting decisions are mostly dependent on the outside ambient condition, potentially causing the mal-defrosting. The developed demand-based defrosting methods are usually unstable or costly for implementation. This paper presents a novel defrosting initiating control strategy based on the convolutional neural network (CNN) for ASHP. The CNN defrosting mechanism is based on the heat pump system's internal operating parameters rather than the outside ambient condition detection. In this paper, a CNN defrosting model is first built and then trained to learn the defrosting logic from an existing time-based method. An experiment of ASHP is conducted to acquire the dataset for CNN training. The CNN is evaluated under 6 different cases, and the maximal and minimal predicted error is 12% and 2% respectively. This demonstrates that CNN is capable to provide accurate frosting predictions and make correct defrosting initiating decisions. It can successfully capture the defrosting logic of time-based method, while avoiding the mal-defrosting caused by time-based method.