基于多任务学习的风速实时预测方法  

Wind speed real-time prediction method based on multi-task learning

作　　者：刘永前[1,2] 周家慷 阎洁[1,2] 韩爽[1,2] 李莉[1,2] Bekhbat Galsan Liu Yongqian;Zhou Jiakang;Yan Jie;Han Shuang;Li Li;Bekhbat Galsan(School of Renewable Energy,North China Electric Power University,Beijing 102206,China;State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources,North China Electric Power University,Beijing 102206,China;Power Engineering School,Mongolian University of Science and Technology,Ulaanbaatar 999097-15141,Mongolia)

机构地区：[1]华北电力大学新能源电力系统国家重点实验室,北京102206 [2]华北电力大学新能源学院,北京102206 [3]蒙古科技大学电力工程学院,蒙古乌兰巴托999097-15141

出　　处：《可再生能源》2021年第4期481-487,共7页Renewable Energy Resources

基　　金：国家自然科学基金项目(U1765201)。

摘　　要：准确的秒级风速实时预测能够提高风电机组的运行状况和控制品质,为电网做出最优调度决策提供辅助信息。目前风速实时预测时间分辨率通常为分钟级,且在小数据集的情况下模型泛化能力弱。文章以时间分辨率为5 s的风速序列为研究对象,提出了基于多任务学习的风速实时预测方法。该方法结合了变分模态分解方法和长短期记忆神经网络。首先,通过变分模态将风速序列分解为一系列信号;然后,建立多任务学习的共享层,使用长短期记忆神经网络提取各分解信号中的共享参数,深度挖掘分享子序列预测任务间的信息;最后,建立多任务学习的特定任务层,借助多个LSTM并行预测分解后的风速子序列,并将多个预测结果叠加得到风速实时预测结果。算例结果表明:所提多任务学习模型在10步、5步预测中的均方根误差总体均值分别为0.80 m/s和0.71 m/s,与经过变分模态分解和未经过变分模态分解的单任务模型预测相比,所提模型均方根误差总体均值在10步预测中分别降低了35.5%和39.8%,在5步预测中分别降低了24.5%和45.8%。Accurate real-time prediction of second-level wind speed can improve the operating conditions and control quality of wind turbines and provide auxiliary information for the power grid to make optimal dispatch decisions.At present,the time resolution of real-time wind speed prediction is usually on the minute level,and the model's generalization ability is weak in the case of small data sets.Therefore,this paper takes the wind speed sequence with a time resolution of 5 s as the research object and proposes a real-time prediction method of wind speed based on multi-task learning,which combines the variational modal decomposition method and long-short term memory neural network.First,the wind speed sequence is decomposed into a series of signals through the variational mode decomposition method.Then,a shared layer for multi-task learning is established,and the long-short term memory neural network is used to extract the shared parameters of each decomposition signal,which digs and shares information between sub-sequence prediction tasks.Finally,the specific task layer of multi-task learning is established,which uses multiple LSTMs to predict the decomposed wind speed sub-sequences in parallel,and the multiple prediction results are superimposed to obtain real-time wind speed prediction results.The practical example results show that the overall mean value of the root mean square error in the 10-step and 5-step prediction of the proposed multi-task learning model is 0.80 m/s and 0.71 m/s.Compared with the single-task model predictions after variational modal decomposition and without variational modal decomposition,the root mean square error of the proposed model is reduced by 35.5%and 39.8%in the 10-step prediction and is reduced by 24.5%and 45.8%in the 5-step prediction.

关 键 词：多任务学习 实时风速多步预测 变分模态分解 长短期记忆网络 

分 类 号：TK81[动力工程及工程热物理—流体机械及工程]