基于CNN-Transformer融合框架的柴油车氨排放预测方法  

Research on diesel vehicle NH_(3) emission prediction method based on CNN-Transformer fusion framework

作　　者：白晓鑫 郭向阳 吴春玲 王凤滨 李旭 刘卫林 BAI Xiao-xin;GUO Xiang-yang;WU Chun-ling;WANG Feng-bin;LI Xu;LIU Wei-lin(CATARC Automotive Test Center(Tianjin)Co.,Ltd.,Tianjin 300300,China;School of Mechanical Engineering,Tianjin University,Tianjin 300072,China)

机构地区：[1]中汽研汽车检验中心(天津)有限公司,天津300300 [2]天津大学机械工程学院,天津300072

出　　处：《中国环境科学》2025年第3期1231-1240,共10页China Environmental Science

基　　金：国家重点研发计划(2022YFC3701800)。

摘　　要：本研究提出了一种基于卷积神经网络(CNN)和Transformer融合框架的柴油车NH_(3)排放预测模型.该模型充分结合了CNN的局部特征提取能力和Transformer的全局依赖关注能力,实现了对实际行驶条件下柴油车NH_(3)排放的高精度预测.研究以一辆N_(3)类柴油车的实际道路排放测试数据为基础,采用Pearson相关系数法进行特征筛选,并利用贝叶斯优化算法对模型关键超参数进行调整,以提升模型性能.此外,应用SHAP算法量化了影响NH_(3)排放的关键因素.结果表明,所提模型在独立数据集上能够高精度预测柴油车实际道路行驶中的NH_(3)浓度排放,其预测的NH_(3)浓度与实际测量值的R^(2)、平均绝对误差(MAE)和均方误差(MSE)分别达到0.986、0.663和2.285,预测性能显著优于传统的随机森林(RF)模型、长短期记忆(LSTM)神经网络模型以及Transformer模型.研究为在用柴油车NH_(3)排放监测提供了一种高效可靠的方法,同时为深入理解影响柴油车实际道路NH_(3)排放的关键因素提供了新的研究思路.In this study,a diesel vehicle NH_(3) emission prediction model based on the fusion framework of Convolutional Neural Network (CNN) and Transformer is proposed.The model was developed by integrating the local feature extraction capability of CNN with the global dependency modeling capability of Transformer,enabling the highly accurate prediction of NH_(3) emissions from diesel vehicles under real road driving conditions.The study was conducted based on the actual on-road emissions test data of an N_(3)-class diesel vehicle.Feature screening was performed using the Pearson correlation coefficient method,and the key hyperparameters of the model were optimized through the application of the Bayesian algorithm,which enhanced its performance.Additionally,the SHapley Additive exPlanations (SHAP) algorithm was utilized to identify the pivotal factors influencing NH_(3) emissions.The results indicated that the proposed model achieved highly accurate predictions of NH_(3) emissions from diesel vehicles in real road driving conditions when tested on an independent dataset.The R^(2),MAE,and MSE values of the predicted NH_(3) concentration compared to the actual measured values were 0.986,0.663,and 2.285,respectively,which were significantly superior to those obtained by the traditional Random Forest (RF) model,the Long Short-Term Memory (LSTM) neural network model,and the Transformer model.This study provided an efficient and reliable method for monitoring NH_(3) emissions from in-use diesel vehicles and offered a novel perspective for elucidating the principal factors influencing NH_(3) emissions from diesel vehicles on the road.

关 键 词：柴油车 排放 NH_(3) 卷积神经网络 TRANSFORMER 

分 类 号：X511[环境科学与工程—环境工程]