基于自监督预训练与时序注意力机制的变压器顶层油温预测  

Transformer Top Oil Temperature Prediction Based on Self-supervised Pre-training and Time-series Attention Mechanism

作　　者：李启明 李彬[1] 刘浩 甘津瑞 石富岭 卢卫疆 杨春萍[1] LI Qiming;LI Bin;LIU Hao;GAN Jinrui;SHI Fuling;LU Weijiang;YANG Chunping(School of Electric and Electronic Engineering,North China Electric Power University,Changping District,Beijing 102206,China;State Grid Smart Grid Research Institute,Changping District,Beijing 102209,China)

机构地区：[1]华北电力大学电气与电子工程学院,北京市昌平区102206 [2]国网智能电网研究院有限公司,北京市昌平区102209

出　　处：《中国电机工程学报》2024年第S01期318-331,共14页PROCEEDINGS OF THE CHINESE SOCIETY FOR ELECTRICAL ENGINEERING

摘　　要：变压器故障预警依赖于更精确可信的变压器顶层油温预测。应用自监督预训练方法将预测模型训练过程中的油温预测任务转变为油温重建任务,泛化模型训练方式的同时提升模型对历史油温态势信息的抽取能力。该文提出双通道预训练时序注意力网络(dual-channel pre-trained time-series attention network,DPAnet)模型,模型是采用时序注意力机制和趋势周期分支的深度神经网络,分别针对油温数据的趋势规律和周期规律实现单时间步级别的建模,从而加强在多时间步上的预测能力。算例分析表明,在以小时为颗粒度的1~72 h短期预测场景下,该文所提出的顶层油温预测模型平均预测损失为1.847,平均确定性系数为0.862,相比其他模型提升预测精度,且具有较强的泛化能力和鲁棒性,有效支撑变压器顶层油温变化趋势分析。Transformer fault early warning relies on more accurate and reliable prediction of transformer top oil temperature.The self-supervised pre-training method is used to transform the oil temperature prediction task in the forecasting model training process into the oil temperature reconstruction task,which generalizes the model training method and improves the model's ability to extract historical oil temperature situation information.A dual-channel pre-trained time-series attention network(DPAnet)model is proposed.The model uses a time-series attention mechanism and a deep neural network with trend cycle branches,respectively targeting the trend laws of oil temperature data.The periodic law realizes modeling at the single time step level,thus enhancing the predictive ability on multiple time steps.The example analysis shows that in the short-term prediction scenario of 1~72 h with hourly granularity,the average prediction loss of the top oil temperature prediction model proposed in this paper is 1.847,and the average certainty coefficient is 0.862,which improves the prediction compared with other models.The model has strong generalization ability and robustness,and effectively supports the analysis of the change trend of oil temperature on the top layer of the transformer.

关 键 词：顶层油温 自监督预训练 时序注意力 深度神经网络 

分 类 号：TM411[电气工程—电器]