In order to effectively avoid the economic loss and safety problems caused by the failure of insulated gate bipolar transistors, the prediction of their parameters can achieve this purpose. In this paper, the switching time, switching loss and collector-emitter turn-off spike voltage are selected as the prediction parameters of insulated-gate bipolar transistors by double-pulse experiments and aging acceleration experiments, and the parameter prediction algorithm of insulated-gate bipolar transistors based on LSTM neural networks is designed by optimizing and pruning operations of ordinary LSTM neural networks. The experimental data of double-pulses and NASA PCoE research center are used to conduct experiments in TensorFlow platform, and the experimental results show that the accuracy of the optimized LSTM network algorithm in this paper improves about 0.85% compared with the ordinary LSTM network algorithm, and about 0.16% compared with the LSTM network algorithm without GRU unit, which provides a new method for IGBT health monitoring and life prediction provides a new model.