A training pattern recognition algorithm based on weight clustering for improving cooling load prediction accuracy of HVAC system. (15th July 2022)

• Record Type: Journal Article
• Authors: Chen, Sihao; Wang, Liangzhu (Leon); Li, Jing; Zhou, Guang; Zhou, Xiaoqing
• Is Part Of: Journal of building engineering. Volume 52(2022)
• Publication Date: 2022-07-15
• Subjects: Cooling load prediction -- Pattern recognition -- Clustering algorithm -- Data preprocessing -- Mode identification; Building -- Periodicals; 690.05
• DOI: 10.1016/j.jobe.2022.104445
• Languages: English
• ISSNs: 2352-7102

Abstract

Abstract: The cooling load-based optimal control is an advanced technology for the efficient operation of heating, ventilation, and air conditioning (HVAC). Thus, the prediction reliability of cooling load plays a key role in HVAC's optimal control. Current publications primarily focused on the structure optimization of prediction models, while less on the clustering-based cooling load prediction. However, the data quality determines the upper limit of the model's prediction performance. Thus, a training pattern recognition algorithm based on weight clustering is proposed for improving cooling load prediction accuracy. Compared with the existing clustering-based prediction methods, the main innovations of the proposed method are: (i) considering the input variables' weights on cooling load in the clustering process; and (ii) investigating the matching between the various prediction models and the K -means clustering algorithm. The case studies showed that the proposed method achieves a significant improvement in the prediction performance, such as MAPEs of the MLR, MNR, and ANN decrease by 34.67%, 35.56%, and 14.53% on average, respectively. Compared with the non-weights clustering method, the introduction of the weights can further improve the above models' prediction accuracy, such as their MAPEs decrease by 6.30%, 7.59%, and 3.07% on average, respectively. These results also demonstrated that the clustering-based prediction method is more suitable for the regression models (e.g., MLR and MNR) with low complexity compared to the ANN. When the clustering number is about 4, the models' prediction performances were more robust. Applying the proposed method to the time-series models (i.e., AR, ARX, and ANN) resulted in their MAPEs as low as 1.79%, 1.78%, and 2.06%, respectively. the proposed method can provide a new idea for improving the accuracy of cooling load prediction. Highlights: The proposed training pattern recognition algorithm has a significant accuracy improvement in the cooling load prediction. Weights of input variables on cooling load were introduced into the clustering and better clustering effects were obtained. The matching between the clustering algorithm and the various prediction models was investigated. The effects of different clustering numbers on the models' performance were studied and the optimal numbers were obtained.