基于在线学习的柴油机起动过程NO_(x)排放预测模型  

Prediction model for NO_(x) emission of starting process of the diesel engine based on online learning

作　　者：杜征宇 李华杰 郭志坤 梁永森 石磊[1] DU Zhengyu;LI Huajie;GUO Zhikun;LIANG Yongsen;SHI Lei(Key Laboratory for Power Machinery and Engineering of Ministry of Education,Shanghai Jiao Tong University,Shanghai 200240,China;92118 Troops of PLA,Zhoushan 316000,China;China North Engine Research Institute,Tianjin 300400,China)

机构地区：[1]上海交通大学动力机械及工程教育部重点试验室,上海200240 [2]中国人民解放军92118部队,浙江舟山316000 [3]中国北方发动机研究所,天津300400

出　　处：《内燃机与动力装置》2024年第5期8-16,共9页Internal Combustion Engine & Powerplant

基　　金：基础产品创新计划(2023027)。

摘　　要：为了构建准确的柴油机起动过程NO_(x)排放预测模型,分析起动过程瞬态运行特点,基于非线性自回归神经网络,结合起动试验数据选取模型特征,对比自注意力机制模型与标准反向传播神经网络模型NO_(x)排放预测效果,发现自注意力机制模型预测效果良好,且均方根误差较小;对比自注意力机制模型在在线梯度下降算法和FTRL算法2种在线学习的预测效果,发现采用FTRL算法可以使模型在未知工况下获得更好的预测性能;对比自注意力机制模型学习前、后的预测效果,发现学习后的模型均方根误差增大,但预测结果能力良好;对比在线学习模型与全数据训练集离线学习模型预测效果,发现两者的决定系数与均方根误差相差不大,预测效果都较好,但在线学习所用数据量和训练时间分别减少了68.7%和73.6%,大幅降低了存储和训练成本。采用FTRL算法的自注意力机制模型可减少数据存储和训练资源成本,并实时反馈预测需求。In order to build an accurate NO_(x)emission prediction model for diesel engine starting process,the transient operation characteristics of the starting process is analyzed,a nonlinear autoregressive neural network is used to select model features based on starting test data.The prediction performance of the self-attention mechanism model and the standard backpropagation neural network model for NO_(x)emissions is compared,it is found that the self-attention mechanism model has good prediction performance and small root mean square error.Comparing the predictive performance of self attention mechanism models in two online learning methods,namely online gradient descent algorithm and FTRL algorithm,it is found that adopting FTRL algorithm could improve the predictive performance of the model under unknown working conditions.Comparing the predictive performance of the self attention mechanism model before and after learning,it is found that the root mean square error of the learned model increases,but the predictive ability is good.Comparing the prediction performance of online learning models and full data training set offline learning models,it is found that the coefficient of determination and root mean square error of both models are not significantly different,and the prediction performance is good.However,the amount of data and training time used in online learning have been respectively reduced by 68.7%and 73.6%,significantly reducing storage and training costs.The results indicate that the self-attention mechanism model using FTRL algorithm could reduce data storage and training resource costs,and provide real-time feedback on prediction requirements.

关 键 词：柴油机 起动过程 NO_(x)预测 自注意力机制 在线学习 

分 类 号：TK421[动力工程及工程热物理—动力机械及工程]