Conformal asymmetric multi-quantile generative transformer for day-ahead wind power interval prediction. (1st March 2023)

• Record Type: Journal Article
• Authors: Wang, Wei; Feng, Bin; Huang, Gang; Guo, Chuangxin; Liao, Wenlong; Chen, Zhe
• Is Part Of: Applied energy. Volume 333(2023)
• Publication Date: 2023-03-01
• Subjects: Wind power -- Interval prediction -- Generative transformer -- Asymmetric interval -- Conformal quantile; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2022.120634
• Languages: English
• ISSNs: 0306-2619

Abstract

Abstract: With the rapid increase in the installed capacity of wind power, day-ahead wind power interval prediction is becoming more and more important. To solve such a challenging problem and provide intervals of higher quality, this paper proposes a prediction method based on conformal asymmetric multi-quantile generative transformer. Herein, the multi-quantile generative transformer is a deep learning model that generates multi-quantile forecast results for next day through one forward propagation process. Then, two quantiles, whose width is the smallest while satisfying the nominal confidence constraint, are selected from the predicted sequence as the upper bound and lower bound of the asymmetric interval. Furthermore, we introduce the conformal quantile regression to calibrate the bounds of the prediction interval to ensure that its coverage rate is as close as nominal confidence. The experiments show that the proposed method surpasses the benchmarks by providing narrower prediction intervals with more accurate empirical coverage probability. Under nominal confidence 90%, it gives prediction intervals with average empirical coverage probability of 90.50% and normalized average width of 0.44 on four wind farms. Compared with symmetric prediction intervals given by common benchmark quantile long short term memory network, the average width is reduced by 19.6%. Highlights: Generative transformer is better adapted to process long sequence time series. Asymmetric interval prediction can significantly reduce the width of PIs. Conformal calibration improves the accuracy of empirical coverage probability of PIs.