Missing wind speed data reconstruction with improved context encoder network. (November 2022)

• Record Type: Journal Article
• Authors: Jing, Bo; Pei, Yan; Qian, Zheng; Wang, Anqi; Zhu, Siyu; An, Jiayi
• Is Part Of: Energy reports. Volume 8(2022)
• Page Start: 3386
• Page End: 3394
• Publication Date: 2022-11
• Subjects: Wind energy -- Data reconstruction -- Missing wind speed data -- Context encoders; Power resources -- Periodicals; Energy industries -- Periodicals; Power resources; Periodicals; Electronic journals; 621.04205
• DOI: 10.1016/j.egyr.2022.02.177
• Languages: English
• ISSNs: 2352-4847

Abstract

Abstract: Missing wind speed data are mainly caused by harsh weather, wind turbine failures, and data transmission errors, which have adverse effects on the performance of wind power forecasting, power curve modeling, and energy assessment. Inspired by context encoders (CE), this paper proposes an improved context encoder network (ICE) for missing wind speed data reconstruction. An auto-encoder architecture with multiple one-dimensional convolutional layers is established for data generation. During network training, a joint loss function that includes reconstruction loss and adversarial loss is presented to obtain the stable and near-real reconstructed wind speed data. We add an Inception layer to the generator network to automatically select the appropriate convolutional filters and then recalibrate the channel relationship between feature maps via the squeeze-and-excitation network. At last, this paper uses wind speed data collected from an on-shore wind farm to verify the effectiveness of the proposed network. The results show that the mean absolute error (MAE) and root mean square error (RMSE) of the ICE network in different data missing rates are 0.019–0.021 and 0.021–0.025, respectively. It has the lowest reconstruction errors compared with six typical data reconstruction methods. Highlights: An improved Context Encoder network is proposed for missing wind speed data reconstruction. An Inception layer optimized by the squeeze-and-excitation network is used for feature extraction. A joint loss that includes reconstruction loss and adversarial loss is employed for network training.