摘要
现有的时间序列任务往往训练一个端到端的模型,由于对标签的重度依赖且对相关仪器要求严苛,此过程往往需要消耗大量人力物力。近年在时间序列的预训练领域中,表示学习特别是对比学习逐渐成为了一个新热点并在各类场景取得了广泛且具有突破性的成效,这些方法通过建模时间序列内在联系给传统时间序列任务带来了提升性能和减少标签重度依赖的可能性。从电力时间序列数据出发,研究表征学习在电力时间序列数据的强大表现并提出一种建模用户粒度的电力时序数据并在该层面上求同存异地提取特征的模型。该模型通过建模社区内用户之间潜在的关系并通过判别式任务增强用户之间的可区分度,使表示学习充分体现用户间的同质元素和异质元素。实验发现本模型在国家电网数据集及其下游任务中得到了优于基线模型10%~15%的效果,并在一些公开数据集中得到一定程度的提升。这些实验都表明了所提出的通过建模用户间关系优化时间序列表示学习模型的优越性。
Existing work on time series often requires training an end-to-end model,a process that is often labor-intensive.In recent years,represent learning of time series is an emerging research area and has achieved breakthroughs in a wide variety of scenarios.These methods can improve the performance of traditional time series tasks and reduce the heavy dependence of labels by modeling the intrinsic relationship of time series.The powerful performance of representation learning on power time series data is studied,and a model is proposed,which models power time series data at the user granularity and extracts features at that level by seeking commonalities and differences.By modeling the potential relationship between users in the community and enhancing the discriminative degree of users through discriminative tasks,the model fully reflects the homogeneous elements and heterogeneous elements among users in the representation learning.The experimental results show that the proposed model is 10%-15%better than the basic model in the national grid dataset and its downstream tasks,and has been improved in some public datasets to a certain extent,demonstrating the superiority of the proposed model.
作者
黄建平
张建松
张旭东
沈思琪
谢裕清
HUANG Jianping;ZHANG Jiansong;ZHANG Xudong;SHEN Siqi;XIE Yuqing(State Grid Zhejiang Power Co.,Ltd.,Hangzhou Zhejiang 310000,China;Information and Communication Branch,State Grid Zhejiang Power Co.,Ltd.,Hangzhou Zhejiang 310000,China)
出处
《电子器件》
CAS
2024年第5期1317-1323,共7页
Chinese Journal of Electron Devices
基金
国网浙江省电力有限公司科技项目(5211XT20008S)。
关键词
时间序列
预训练
电力数据
特征提取
表示学习
time series
pre-train
electricitydata
feature extraction
represent learning
作者简介
黄建平(1972-),男,浙江杭州人,硕士,高级工程师,研究方向:从事数据治理、大数据等,huang_jianping@zj.sgcc.com.cn。
