基于CM⁃BiLSTM 的APU 排气温度预测模型

APU exhaust gas temperature prediction model based on CM⁃BiLSTM

排气温度(EGT)是表征辅助动力装置(APU)性能的重要指标,针对EGT长周期预测精度低、传统线性预测方法难预测等问题,以长短期记忆网络(LSTM)模型为基础,提出CM⁃BiLSTM模型对APU未来时刻EGT进行预测。首先,采用基于相关系数矩阵(CM)的方法对APU主要性能参数进行特征选择,提取出与EGT相关性最高的参数,即转速,并以此构建训练集、测试集和验证集;然后,将转速和EGT数据归一化处理后输入CM⁃BiLSTM网络进行训练;最后,将验证集输入模型进行预测,预测结果经过反归一化后输出最终EGT的预测值。实验结果表明,CM⁃BiLSTM模型输出的EGT预测值能够较好地逼近真实值,CM⁃BiLSTM模型预测结果的平均绝对误差与均方根误差均低于LSTM和CM⁃LSTM模型,有效提高了预测精度。

The exhaust gas temperature(EGT)is an important indicator to characterize the performance of auxiliary power units(APU).In allusion to the problems such as the low accuracy of long⁃periodic prediction of EGT and the difficulty of traditional linear prediction methods,a CM⁃BiLSTM model is proposed to predict the EGT of APU at a future moment by taking the long short term memory(LSTM)network model as the basis.The method based on the correlation coefficient matrix(CM)is used to perform the feature selection for the APU main performance parameters,and extract the rotation speed,which has the highest correlation with EGT.On this basis,the training set,test set and verification set were constructed.The rotation speed and EGT data are normalized and then input into the CM⁃BiLSTM network for training.The validation set is input to the model for the prediction,and the final EGT prediction value is output after the inverse normalization of the prediction results.The experimental results show that the EGT prediction value output by the CM⁃BiLSTM model can better approximate the true value,the average absolute error and root⁃mean⁃square error of the prediction results of the CM⁃BiLSTM model are lower than those of the LSTM and CM⁃LSTM models,which effectively improves the prediction accuracy.