| 引用本文: | 张越,臧海祥,程礼临,刘璟璇,卫志农,孙国强.基于自适应时序表征和多级注意力的超短期风电功率预测[J].电力自动化设备,2024,44(2):117-125. |
| ZHANG Yue,ZANG Haixiang,CHENG Lilin,LIU Jingxuan,WEI Zhinong,SUN Guoqiang.Ultra-short-term wind power forecasting based on adaptive time series representation and multi-level attention[J].Electric Power Automation Equipment,2024,44(2):117-125. |
|
| 摘要: |
| 针对风电功率数据包含的多尺度时间信息难以描述、现有方法未充分考虑气象因素对于风电功率动态耦合的影响而导致的预测性能下降等问题，提出了一种基于自适应时序表征和多级注意力的超短期风电功率预测方法。采用时序嵌入层对风电功率序列进行表征以获取其周期、非周期模式，并引入自注意力捕捉高维风电功率序列的自相关性；利用交叉注意力重构风电功率与气象因素，形成包含两者耦合关系的多维特征序列；利用一维卷积神经网络沿时间、特征方向分别挖掘多维特征序列的时间相关性和空间相关性，进而利用长短期记忆网络提取相应的时序特征，并将所得时序特征经全局注意力去噪和门控机制融合后输入全连接层，分别进行点预测和区间预测。实验结果表明，所提方法能够获得准确的点预测值和可靠的预测区间。 |
| 关键词: 风电功率 超短期预测 多级注意力 深度学习 时空相关性 点预测 区间预测 |
| DOI：10.16081/j.epae.202306018 |
| 分类号: |
| 基金项目:国家自然科学基金资助项目(52077062) |
|
| Ultra-short-term wind power forecasting based on adaptive time series representation and multi-level attention |
|
ZHANG Yue, ZANG Haixiang, CHENG Lilin, LIU Jingxuan, WEI Zhinong, SUN Guoqiang
|
|
College of Energy and Electrical Engineering, Hohai University, Nanjing 211100, China
|
| Abstract: |
| In view of the problems that the multi-scale time information in wind power data is difficult to describe and the existing methods fail to fully consider the dynamic coupling effects of meteorological factors on wind power, which lead to forecasting performance degradation, an ultra-short-term wind power forecasting method based on adaptive time series representation and multi-level attention is proposed. The wind power sequence is represented by the time embedding layers to obtain its periodic and non-periodic patterns, and then the self-attention is introduced to capture the autocorrelation of the high-dimensional wind power sequence. The wind power and meteorological factors are reconstructed by cross attention to form a multi-dimensional feature sequence containing the coupling relationships between them. One-dimensional convolutional neural networks are used to mine both temporal correlation and spatial correlation of the multi-dimensional feature sequences along the time direction and feature direction respectively. Then, the corresponding temporal features are extracted by the long short-term memory network, and the obtained temporal features are denoised by the global attention and integrated by the gating mechanism, which are then fed to a fully-connected layer to generate point forecasting result and interval forecasting result respectively. Experimental results show that the proposed method can obtain both accurate point forecasting results and reliable forecasting intervals. |
| Key words: wind power ultra-short-term forecasting multi-level attention deep learning spatiotemporal correlation point forecasting interval forecasting |
