ZhejiangElectricPower第42卷第1期2023年1月Vol.42，No.01Jan.25.2023Transformer在水电机组异常指标预测的应用林烨敏1，王宁1，邱荣杰1，汤宇超1，周冠群2，李泽洲2，王中亚2（1.国网浙江省电力有限公司紧水滩水力发电厂，浙江丽水323000；2.宁波工业互联网研究院，浙江宁波315016）摘要：水电站机组日常检修、维护及异常检测的工作量巨大，传统人工监测的工作方式容易导致异常问题被遗漏或误判，采用深度学习算法对数据建模并监测异常情况可以降低成本，提升安全可靠性。结合Transformer网络对长期序列高效准确建模的能力以及GAN（生成对抗网络）架构的数据生成训练策略，利用TransGAN模型对水电机组监测数据进行生成式建模，并主动发现异常数据点。TransGAN模型在水电站机组实测中达到了97.76%的查准率和99.23%的查全率，异常点检出延迟低于0.1s，实现了实时高精度异常监控功能。关键词：水电机组；异常监测；数据降维；Transformer；生成对抗网络DOI:10.19585/j.zjdl.202301014开放科学（资源服务）标识码（OSID）：ApplicationofTransformerinanomalyindicatorsforecastingofhydropowerunitsLINYemin1，WANGNing1，QIURongjie1，TANGYuchao1，ZHOUGuanqun2，LIZezhou2，WANGZhongya2（1.JinshuitanHydropowerPlantofStateGridZhejiangElectricPowerCo.，Ltd.，Lishui，Zhejiang323000，China；2.NingboIndustrialInternetInstitute，Ningbo，Zhejiang315016，China）Abstract:Theworkloadsofroutinerepair，maintenance，andabnormalitydetectionofhydropowerunitsareheavy.Therefore，traditionalmanualmonitoringmayleaveoutormisjudgeabnormalities.Deeplearningalgorithmsareusedfordatamodelingandmonitoringabnormalitiestoreducecostsandimprovesafetyandreliability.WiththehelpoftheTransformerneuralnetworks，theefficientandaccuratemodelingcapacityoflong-termsequencesandtheGAN（generativeadversarialnetwork）architecturedataareusedtogenerateatrainingstrategy.ATransGANmodelisusedforgenerativemodelingofthemeasureddataofhydropowerunitsandproactivelydetectsabnormaldatapoints.TheTransGANmodelachievesadetectionaccuracyrateof97.76%andarecallrateof99.23%inhydro⁃powerdatameasurement.Theanomalydetectiondelayislessthan0.1s.Thereal-timehigh-precisionanomalymoni⁃toringfunctionisrealized.Keywords:h...
