基于LSTM神经网络的柴油机NOx排放预测  

Prediction of Diesel Engine NOx Emissions Based on Long-Short Term Memory Neural Network

作　　者：戴金池 庞海龙 俞妍 卜建国 资新运 Dai Jinchi;Pang Hailong;Yu Yan;Bu Jianguo;Zi Xinyun(Postgraduate Team,Army Military Transportation University,Tianjin 300161,China;Military Vehicle Engineering Department,Army Military Transportation University,Tianjin 300161,China)

机构地区：[1]陆军军事交通学院研究生队,天津300161 [2]陆军军事交通学院军用车辆工程系,天津300161

出　　处：《内燃机学报》2020年第5期457-463,共7页Transactions of Csice

基　　金：国家重点研发计划资助项目(2017YFC0211104).

摘　　要：柴油机NOx排放是机动车排放污染物的主要来源,有效的NOx排放预测模型是选择性催化还原技术(SCR)控制和车载诊断系统(OBD)完成SCR监测的基础.利用长短期记忆(LSTM)神经网络预测某柴油机的NOx排放,LSTM神经网络能够记忆时间序列先前的输入并将其用于当前的预测.将稳态工况与瞬态工况整合成新的混合工况,并在划分的测试集和全球统一瞬态试验循环(WHTC)工况上验证模型精度,结果表明:LSTM神经网络模型能够同时在稳态过程与瞬态过程取得较高的预测精度和稳定性,整合工况测试集的预测误差均方根为55.33×10-6,并且具备较强的泛化能力.Diesel engine is the main source of NOx emissions in motor vehicles.An effective NOx emission prediction model takes an important role in control and on-board diagnosis(OBD)of the selective catalytic reduction(SCR)in order to reduce NOx emissions.The long-short term memory neural network(LSTM NN),which can remember previous inputs in time series and then use them for current input predictions,was used to predict NOx emissions of a diesel engine.The prediction accuracy of the LSTM NN model was verified by various tests including the steady cycle,the transient cycle,the integrated cycle from the above cycles and the world harmonized transient cycle(WHTC)as well.The results show that the LSTM NN model has a good generalization ability.It can give a high prediction accuracy and a good data stability in both steady and transient operating processes,in which the root mean square error is 55.33×10^-6.

关 键 词：柴油机 NOX排放 预测模型 长短期记忆(LSTM)神经网络 

分 类 号：TK421.5[动力工程及工程热物理—动力机械及工程]