Refrigerant charge fault detection method of air source heat pump system using convolutional neural network for energy saving

被引:102
作者
Eom, Yong Hwan [1 ]
Yoo, Jin Woo [1 ]
Hong, Sung Bin [1 ]
Kim, Min Soo [1 ]
机构
[1] Seoul Natl Univ, Dept Mech & Aerosp Engn, Seoul 08826, South Korea
基金
新加坡国家研究基金会;
关键词
Heat pump system; Refrigerant charge fault detection; Convolutional neural network; Quantitative prediction; Classification; Regression; CO2/PROPANE MIXTURES; COOLING PERFORMANCE; CONDITIONING SYSTEM; VRF SYSTEM; DIAGNOSIS; STRATEGY;
D O I
10.1016/j.energy.2019.115877
中图分类号
O414.1 [热力学];
学科分类号
摘要
In heat pumps, refrigerant leakage is one of the frequent faults. Since the systems have the best performance at the optimal charge, it is essential to predict refrigerant charge amount. Hence, the refrigerant charge fault detection (RCFD) methods have been developed by researchers. Due to improvements in computing speed and big-data, data-driven techniques such as artificial neural networks (ANNs) have been highlighted recently. However, most existing ANN-based RCFD methods use low-performance shallow neural networks (SNNs) and require the features extracted by experts' experiences. Also, they have some critical limitations. First, they cannot provide quantitative information on recharge amount due to a simple classification such as undercharge or overcharge. Second, many ANN-based RCFD methods can be used in one operation mode (cooling or heating mode). To improve the limitations, a novel RCFD strategy based on convolutional neural networks (CNNs) was suggested. Two prediction models using classification and regression can predict the quantitative refrigerant amount in both cooling and heating mode with a single model. The mean accuracy of the CNN-based classification model was 99.9% for the learned cases. Also, the CNN-based regression model showed the excellent prediction performance with root-mean-square (RMS) error of 3.1% including the untrained refrigerant charge amount data. (C) 2019 Published by Elsevier Ltd.
引用
收藏
页数:13
相关论文
共 45 条
[1]  
[Anonymous], THESIS
[2]  
[Anonymous], B210 AMCA INT INC
[3]  
[Anonymous], 2009, 2009 IEEE 12 INT C C
[4]  
[Anonymous], 2012, NEURAL NETWORKS TRIC, DOI DOI 10.1007/978-3-642-35289-8_32
[5]  
[Anonymous], 2016, ADAPT COMPUT MACH LE
[6]  
[Anonymous], 1230 ANSIAHRI
[7]  
[Anonymous], NIST SPECIAL PUBLICA
[8]  
[Anonymous], 2001, ELEMENTS STAT LEARNI
[9]  
[Anonymous], P IEEE INT C NEUR NE
[10]  
[Anonymous], NATL C BUILDING COMM