Refrigerant charge fault diagnosis strategy for VRF systems based on stacking ensemble learning. (15th April 2023)

• Record Type: Journal Article
• Authors: Zhang, Li; Cheng, Yahao; Zhang, Jianxin; Chen, Huanxin; Cheng, Hengda; Gou, Wei
• Is Part Of: Building and environment. Volume 234(2023)
• Publication Date: 2023-04-15
• Subjects: Variable refrigerant flow system -- Refrigerant charge fault -- Stacking ensemble learning -- Fault diagnosis -- Recursive feature elimination; Buildings -- Environmental engineering -- Periodicals; Building -- Research -- Periodicals; Constructions -- Technique de l'environnement -- Périodiques; Electronic journals; 696
• DOI: 10.1016/j.buildenv.2023.110209
• Languages: English
• ISSNs: 0360-1323

Abstract

Abstract: The VRF system frequently has refrigerant charge amount (RCA) fault, and this causes a large amount of building energy waste. The research on data-driven models used to diagnose this fault mostly focuses on the optimization of a single model, and it is difficult to maintain good performance at different fault levels. In view of this, this paper proposes the RCA fault diagnosis strategy of VRF system based on Stacking ensemble learning. Firstly, the strategy selects the low dimensional feature set through Recursive Feature Elimination (RFE) and correlation analysis. Then the initial Stacking ensemble learning model consists of two levels of learners. The output of the first-level learners and the original feature set are used as the feature input of the second-level learners. Then, the composition of the first-level learners in the model is adjusted according to the feature importance ranking results of the RFE method. The results show that the classification accuracy (CA) of the model optimized by the proposed strategy in the training set and the test set is improved by 3.9% and 4.02% respectively, and there is little difference between the two, indicating the model generalization ability is improved. Highlights: The proposed strategy is based on the stacking ensemble learning method. The experimental data used covers ten levels of refrigerant charge fault. Feature selection process is added to explain model adjustment process. The diagnosis performance of five simple models and the adjusted model is discussed.