[Poster Presentation]Domain extended attention network for ultra-short-term wind power prediction

Domain extended attention network for ultra-short-term wind power prediction

ID：364 View Protection：ATTENDEE Updated Time：2024-05-27 08:58:34 Hits：514 Poster Presentation

Start Time：2024-05-30 15:40 （Asia/Shanghai）

Duration：20min

Session：[S5] Smart Energy and Clean Power Technology » [S5-1] Afternoon of May 30th

No files

Abstract

Ultra-short-term wind power prediction plays a key role in countless fields, from power forecasting and environmental science to industrial processes . Capturing ultra-short-term dependency and seasonal patterns in time series data is an important issue to improve the accuracy of time series data prediction. We propose Domain Extended Attention Network (DEAN) to effectively capture the characteristics of ultra-short-term dependence through reducing redundancy and expanding receptive field, namely Domain Extended Attention, while maintaining high attention to maintaining computing efficiency and reducing the impact of redundancy. In this paper, by comparing the DEAN model with other six different baseline models on four different prediction lengths, the prediction performance of DEAN model on more than three prediction lengths has reached the current SOTA level, especially on the 96 and 720 prediction lengths, the performance of DEN model is 14.29% and 5.28% higher than that of the second best model respectively.

Keywords

ultra-short-term forecasting; wind power prediction; attention, redundancy; receptive field.

Speaker

Likang Lin

South China Agricultural University