收稿日期:2021-04-01修回日期:2021-04-27第40卷第2期计算机仿真2023年2月文章编号:1006-9348(2023)02-0123-05基于LSTM网络的汽轮机转子热应力预测方法杨晨,柴京,张涛,王晓升(重庆大学能源与动力工程学院,重庆400044)摘要:在电厂灵活运行期间,转子内部因温度梯度较大而产生热应力,导致转子疲劳损伤。而传统有限元分析热应力的方法无法满足实时监测的需求。研究了一种基于数据驱动的LSTM神经网络模型。模型具有从历史序列数据中学习深度信息的能力。通过多组超参数对比实验,发现在神经元数量6,单元节点28,学习率0.005,Dropout比例0.5时网络预测效果较好;在冷态启动过程下使用LSTM神经网络模型的热应力预测数据与有限元样本数据相比,RMSE为7.8740MPa,最大热应力误差9.7480MPa。结果表明,上述模型相比传统有限元计算时间大大缩短,在保证较高精度同时,也能够满足未来实时在线监测的需要。关键词:汽轮机转子;热应力;长短期记忆;超参数中图分类号:TP183文献标识码:BPredictionMethodofSteamTurbineRotorThermalStressBasedonLong-Short-TermMemoryNetworkYANGChen,CHAIJing,ZHANGTao,WANGXiao-sheng(SchoolofEnergyandPowerEngineering,ChongQingUniversity,Chongqing400044,China)ABSTRACT:Duringtheflexibleoperationofthepowerplant,thermalstressisgeneratedintherotorduetothelargetemperaturegradient,whichleadstothefatiguedamageoftherotor.Thispaperstudiesdadata-drivenLSTMneuralnetworkmodel,whichcanautomaticallylearnfeaturesfromhistoricaldata.Throughanumberofsuperparametricex-periments,itiswasfoundthatthenetworkpredictioneffectisbetterwhenthenumberofneuronsis6,thenumberofunitnodesis28,thelearningrateis0.005,andthedropoutratiois0.5.Comparedwiththefiniteelementsampleda-ta,theRMSEofLSTMneuralnetworkmodelis7.8740mpa,andthemaximumthermalstresserroris9.7480mpa.Theresultsshowthatthecalculationtimeofthemodelisgreatlyshortenedcomparedwiththetraditionalfiniteele-mentmethod,whichcanensurehighaccuracyandmeettheneedsofreal-timeonlinemonitoringinthefuture.KEYWORDS:Steamturbinerotor;Thermalstress;Longandshort-termmemory;Hyperparameter1引言随着可再生能源的发展,传统的燃煤电厂将承担更多的调峰任务,这对电厂的灵活运行提出了更高...