Black swan event small-sample transfer learning (BEST-L) and its case study on electrical power prediction in COVID-19. (1st March 2022)

• Record Type: Journal Article
• Authors: Hu, Chenxi; Zhang, Jun; Yuan, Hongxia; Gao, Tianlu; Jiang, Huaiguang; Yan, Jing; Wenzhong Gao, David; Wang, Fei-Yue
• Is Part Of: Applied energy. Volume 309(2022)
• Publication Date: 2022-03-01
• Subjects: Transfer learning -- Black swan event -- Small-sample learning -- COVID-19 -- Load forecasting; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2021.118458
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: A transfer learning framework and a CNN model are proposed for the black swan small-sample events. (BEST-L). Experiments show that the BEST-L outperforms traditional methods. The transfer learning method can effectively use small samples to improve the model's generalization. performance. The BEST-L is tested for the load forecasting case with the COVID-19 scenario in Central China. Abstract: The black swan event will usually cause a great impact on the normal operation of society. The scarcity of such events leads to a lack of relevant data and challenges in dealing with related problems. Different situations also make the traditional methods invalid. In this paper, a transfer learning framework and a convolutional neuron network are proposed to deal with the black swan small-sample events (BEST-L). Taking the COVID-19 as a typical black swan event, the BEST-L is utilized to achieve accurate mid-term load forecasting using the relationship between economy and electricity consumption. The experiment results show that the transfer learning model can effectively learn the basic knowledge about the relationship between the adopted input and output data and use a relatively small amount of data during the black swan event to improve the target areas' generalization. The approach and results can provide an effective approach to respond and react to sudden changes quickly and effectively in similar open problems.