摘要
针对传统超短期风电功率预测方法难以应对海量强波动性数据,且对时间序列处理能力有限的问题,提出一种基于改进的深度可分离卷积神经网络(the improved depthwise separable convolution neural networks,IDSCNN)、注意力机制(attention mechanism,AM)、长短期记忆神经网络(long short-term memory neural network,LSTM)的超短期风电功率组合预测方法。首先,基于IDSCNN设计能够匹配风电场群时空维度变换的可分离卷积核尺寸,对数值天气预报数据、实测功率数据进行一次时空特征提取,以获取气象–功率时空特征。然后,结合AM强化一次时空特征长时间序列中局部重要信息的贡献程度,筛选出与未来预测功率密切相关的二次时空特征,以作为LSTM预测模型的输入时间序列。最后,建立包含改进的深度可分离卷积层、注意力权重分配层、LSTM预测层的IDSCNN-AM-LSTM组合神经网络超短期风电功率预测模型。仿真结果表明:该方法能够利用深度学习在挖掘高维非线性特征时的优势,对多个风电场之间的时空相关性进行充分学习,而且在单步风场功率预测和多步集群功率预测上,与其他预测模型相比均具有较高的预测精度和较好的时序学习能力。
It is difficult for traditional ultra-short-term wind power prediction methods to deal with the numerous data with highly variation, and these methods have limitations in processing time series. Therefore, an ultra-short-term wind power combined prediction method including the improved depthwise separable convolution neural networks(IDSCNN),attention mechanism(AM), and long short-term memory neural network(LSTM) is proposed. Firstly, the separable convolution kernel size is designed by IDSCNN that can match the spatiotemporal dimension transformation of wind farms,so as to extract the primary spatiotemporal feature from the numerical weather prediction data and the measured power data to obtain key meteorological and power spatiotemporal features. Secondly, the AM is used to strengthen the contribution of local important information in long time series of the primary spatiotemporal feature, and to screen out the secondary spatiotemporal feature which is closely related to the future prediction outcomes. Subsequently, the captured secondary feature is sent to the LSTM prediction model as the input time series. Finally, a new IDSCNN-LSTM network structure is established which is combined with IDSCNN layer, attention weight distribution layer and LSTM prediction layer to realize ultra-short-term wind power prediction. The simulation results show that the proposed method can not only be adopted to sufficiently learn the spatiotemporal dependencies among wind farms assisted by the advantages of deep learning in capturing high dimensional nonlinear features, but also has higher prediction accuracy and better timing learning ability in single-step wind farm prediction and multi-step cluster prediction compared with other models.
作者
李卓
叶林
戴斌华
於益军
罗雅迪
宋旭日
LI Zhuo;YE Lin;DAI Binhua;YU Yijun;LUO Yadi;SONG Xuri(College of Information and Electrical Engineering,China Agricultural University,Beijing 100083,China;China Electric Power Research Institute,Beijing 100192,China)
出处
《高电压技术》
EI
CAS
CSCD
北大核心
2022年第6期2117-2127,共11页
High Voltage Engineering
基金
国家电网公司总部科技项目(5108-202155037A-0-0-00)。
关键词
超短期风电功率预测
深度可分离卷积
注意力机制
长短期记忆神经网络
时间序列
ultra-short-term wind power prediction
depth separable convolution
attention mechanism(AM)
long short-term memory neural network(LSTM)
time-series
