基于VMD-LSTM-SVR的IGBT寿命特征时间序列预测

Time Series Prediction of IGBT Lifetime Feature Based on VMD-LSTM-SVR

绝缘栅双极型晶体管(IGBT)失效是变频器等电力电子设备故障的主要原因,精确预测其寿命是解决该问题的方法之一,这对寿命预测模型的准确性和可靠性提出了更高要求。关断瞬态尖峰电压(Vce,peak)可以反映IGBT的老化状态,首先通过变分模态分解(VMD)技术将Vce,peak构成的时间序列分解为趋势序列和波动序列,再利用长短期记忆(LSTM)网络的时间序列特征提取优势和支持向量机回归(SVR)的非线性求解能力,建立VMD-LSTM-SVR组合模型,提升模型的预测性能。模型预测对比实验结果表明,VMD-LSTM-SVR模型提升了IGBT寿命特征时间序列预测能力,与其他模型相比,该模型的预测精度指标均方根误差下降至0.0411 V,决定系数提升至0.75111。

The failure of insulated gate bipolar transistor(IGBT)is the main reason for the failure of power electronic equipment such as frequency converters.Accurate prediction of its lifetime is one of the methods to address this problem,which puts forward higher demands on the accuracy and reliability of lifetime prediction models.The turn-offtransient peak voltage(Veopek)can reflect the aging state of ICBT.Firstly,the time series formed by V.peak was decomposed into trend and fluctuation series by variational mode decomposition(VMD)technology.Then,by using the time series feature extraction advantage of long short-term memory(LSTM)network and the non-linear solving ability of support vector machine regression(SVR),a VMD-LSTM-SVR combination model was established to improve the predictive performance of the model.The experimental results of model prediction and comparison show that the VMD-LSTM-SVR model improves the time series predictive ability of IGBT lifetime feature.Compared with other models,the root mean square error of the prediction accuracy index of this model decreases to 0.0411 V,and the determination coefficient increases to 0.75111.