Ensemble learning-based approach for residential building heating energy prediction and optimization. (15th May 2023)

• Record Type: Journal Article
• Authors: Zhang, Jianxin; Huang, Yao; Cheng, Hengda; Chen, Huanxin; Xing, Lu; He, Yuxuan
• Is Part Of: Journal of building engineering. Volume 67(2023)
• Publication Date: 2023-05-15
• Subjects: Energy consumption prediction -- Ensemble learning -- Heating station -- Machine learning -- Optimization; Building -- Periodicals; 690.05
• DOI: 10.1016/j.jobe.2023.106051
• Languages: English
• ISSNs: 2352-7102

Abstract

Abstract: Accurate building energy consumption prediction is critical for engineers to design optimized operational strategies for building heating, ventilation, and air-conditioning systems. In this paper, an stacking ensemble learning-based model is established based on the operational data of a district resident buildings heating station for building heating system energy consumption prediction. The ensemble model is optimized by outlier processing, feature selection, parameter optimization based on grid search. A new feature based on Exponentially Weighted Moving Average (EWMA) algorithm was proposed to take historical energy feature into consideration. The performance of the ensemble model and four base machine learning methods, including multiple linear regression, extreme learning machine, extreme gradient boosting and support vector regression, are evaluated. Compared with the four base models, the Mean Absolute Error (MAE) of the ensemble model decreases by 4.36%–71.70%, and the Root Mean Squared Error (RMSE) by 3.80%–49.73%. Using the new feature based on EWMA can further reduce the MAE and RMSE of the ensemble model by 10.36% and 19.89%, respectively. The result proves that the proposed ensemble model with the added historical feature effectively improves the prediction model's accuracy for building heating energy consumption. Highlights: A case study on a heating station system in Beijing was illustrated. A novel heat supply energy consumption prediction strategy was proposed. A stacking ensemble model was trained and verified in this study.