Wind power forecasting using attention-based gated recurrent unit network. (1st April 2020)

• Record Type: Journal Article
• Authors: Niu, Zhewen; Yu, Zeyuan; Tang, Wenhu; Wu, Qinghua; Reformat, Marek
• Is Part Of: Energy. Volume 196(2020)
• Publication Date: 2020-04-01
• Subjects: Wind power forecasting -- Recurrent neural network -- Feature selection -- Attention mechanism -- Deep learning; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2020.117081
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: Wind power forecasting (WPF) plays an increasingly essential role in power system operations. So far, most forecasting models have focused on a single-step-ahead WPF, and the obtained results are insufficient for planning and operations of the power system due to the intermittent and fluctuated nature of wind. At the same time, most of the current multi-step-ahead WPF models neglect the correlation between different forecasting tasks. In this paper, we propose a novel sequence-to-sequence model using the Attention-based Gated Recurrent Unit (AGRU) that improves accuracy of forecasting processes. It embeds the task of correlating different forecasting steps by hidden activations of GRU blocks. In addition, an attention mechanism is designed as a feature selection method to identify the most important input variables. To validate the effectiveness of the proposed AGRU model, three different case studies focused on forecasting accuracy, computational efficiency, and feature selection abilities are carried out. Their performances are compared with various wind power forecasting benchmarks. Highlights: Correlations among different step ahead forecasting tasks improve the forecasting accuracy and stability. The sequence-to-sequence deep learning scheme is deployed for multi-step ahead forecasting. An attention mechanism is designed as the feature selection method. Identification of the most important factors influencing wind power generation processes. The results can provide valuable insights for deep learning-based method investigation.