Twin model: A continuous learning method for short-term online forecasting of office building energy consumption considering extreme scenarios
Fang, Haizhou; Huang, Guocheng; Zhang, Zhenyu; Yuan, Xiaolei; Zhai, Xiaoqiang (2026-10-19)
Katso/ Avaa
Sisältö avataan julkiseksi: 20.10.2028
Post-print / Final draft
Fang, Haizhou
Huang, Guocheng
Zhang, Zhenyu
Yuan, Xiaolei
Zhai, Xiaoqiang
19.10.2026
Energy
366
Elsevier
School of Energy Systems
Kaikki oikeudet pidätetään.
© Elsevier
© Elsevier
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe20261028138055
https://urn.fi/URN:NBN:fi-fe20261028138055
Tiivistelmä
The online forecasting of building energy consumption is essential for the effective functioning of building energy systems and demand response. Nevertheless, the considerable discrepancies in building operational settings and the scarce availability of representative operational data markedly diminish prediction precision, particularly in intricate and harsh operational contexts. The study develops a novel twin model framework, centered on continual learning, employing data enriched by physical simulations to bolster learning, with the objective of enhancing the generalisation capacity of online short-term energy consumption forecasts for office spaces. The thoroughness of training information is assessed according to the distinctiveness, patterns, and daily variations of the training data. When the integrity of the information is diminished, supplementary feature variables will be used to enhance predictive efficacy. The suggested framework assesses the reliability of training data and alternates between data-centric continuous learning and physics-informed data enhancement to rebuild the predictive model when past operational data is lacking. The findings from the case study demonstrate that the suggested twin model significantly enhances the dependability of short-term online forecasts for office buildings, both inside and outside the historical operational conditions. In situations that exceed historical experience, the simulation-based proof of concept demonstrated that the suggested approach diminished the CV-RMSE of adaptive forecasts by more than 70% relative to continuous learning utilizing historical data. The suggested system offers a robust continuous learning approach for dependable short-term online energy forecasting in office environments and facilitates practical uses for forthcoming smart building functionalities and digital twin applications.
Lähdeviite
Haizhou Fang, Guocheng Huang, Zhenyu Zhang, Xiaolei Yuan, Xiaoqiang Zhai. (2026). Twin model: A continuous learning method for short-term online forecasting of office building energy consumption considering extreme scenarios. Energy, Volume 366. DOI: https://doi.org/10.1016/j.energy.2026.142822
Kokoelmat
- Tieteelliset julkaisut [1956]