=================================DOI:10．13290/j．cnki．bdtjs．2023．01．01166半导体技术第48卷第1期2023年1月基于SSA-LSTM模型的IGBT时间序列预测研究冷丽英，付建哲*，宁波(西安中车永电捷通电气有限公司，西安710016)摘要:针对绝缘栅双极型晶体管(IGBT)长工作周期导致的老化失效问题，提出一种基于麻雀搜索算法(SSA)优化长短期记忆(LSTM)网络的IGBT时间序列预测方法。首先分析IGBT疲劳失效的原因，选取某IGBT老化数据集中的集射极峰值电压为失效特征量，进行二次指数滤波以增大数据下降趋势。然后利用Matlab搭建LSTM模型，并采用SSA对网络模型中学习率、隐藏层节点数和训练次数进行寻优以得到最优网络。最后选取常用回归预测性能评估指标对LSTM模型与SSA-LSTM模型预测结果进行对比分析。结果表明，SSA-LSTM模型的预测结果平均绝对误差、均方根误差及平均绝对百分比误差分别降低了0.016%、0.022%和0.202%，证明所提方法预测精度高，可在一定程度上评估IGBT的寿命。关键词:麻雀搜索算法(SSA);长短期记忆(LSTM)网络;绝缘栅双极型晶体管(IGBT);特征参数;时间序列预测中图分类号:TN322.8;TN306文献标识码:A文章编号:1003－353X(2023)01－0066－07ＲesearchonIGBTSequentiallyPredictionBasedonSSA-LSTMModelLengLiying，FuJianzhe*，NingBo(CＲＲCXi＇anYongeJietongElectricCo．，Ltd．，Xi＇an710016，China)Abstract:Aimingattheagingfailureproblemcausedbythelongoperatingcycleofinsulatedgatebipolartransistors(IGBTs)，anIGBTsequentiallypredictionmethodwasproposedwhichoptimizingthelongshort-termmemory(LSTM)networkbasedonthesparrowsearchalgorithm(SSA)．Firstly，thecausesofIGBTfatiguefailurewereanalyzed，andsecondaryexponentialfilteringwasperformedtoincreasethedownwardtrendofthedatabasedonthepeakcollector-emittervoltagesinanIGBTagingdatasetasthefailurecharacteristicquantity．Secondly，theLSTMmodelwasbuiltbyMatlab，andtheSSAwasusedtomajorizethekeyparameterssuchaslearningrate，numberofhiddenlayernodesandnumberoftrainingsinthenetworkmodeltoobtaintheoptimalnetwork．Finally，thecommonlyusedregressionpredictionperformanceevaluationindicatorswereselectedtocompareandanalyzethepredictionresultsoftheLSTMandtheSSA-LSTMmodel．Theresultsshowthatthepredictionmeanabsoluteerror，ro...
