适用发动机性能预测的先进机器学习方法

Advanced machine learning methods for engine performance prediction

为了提高六缸重型柴油-天然气双燃料发动机的性能、减少污染物排放,使用反馈神经网络、支持向量机和高斯处理3种先进的机器学习方法来建立以发动机的转速、扭矩、柴油预喷提前角、柴油预喷比例和天然气替代率作为输入,以NO_(x)、CO_(2)和比油耗作为输出的预测模型,对比3种机器学习模型的预测结果。结果表明:高斯处理(Gaussian processing, GP)模型的预测精度最高,其输出变量的总体决定系数分别为0.960 1、0.991 9和0.993 5,相比于反馈神经网络(feedback neural network, FNN)和支持向量机(support vector machine, SVM),NOx预测精度分别提高3.7%和2.6%,CO_(2)排放预测精度提高约3%,比油耗(brake specific fuel consumption, BSFC)预测精度分别约提高4%和8%;GP模型预测NO_(x)、CO_(2)和BSFC的总体均方相对误差均小于0.13,总体平均绝对百分比误差均小于0.1%。

Three advanced machine learning methods(feedback neural network,support vector machine and Gaussian processing)were used to improve the performance of the six-cylinder heavy-duty diesel natural gas dual fuel engine and reduce pollutant emissions.A prediction model was established with engine speed,torque,diesel preinjection advance angle,diesel preinjection ratio and natural gas replacement rate as inputs and NO_(x),CO_(2) and brake specific fuel consumption as outputs.By comparing the prediction results of the three machine learning models,it is found that GP model has the highest prediction accuracy,and the overall determination coefficients of its output variables are 0.9601,0.9919 and 0.9935,respectively.Compared with FNN and SVM,the prediction accuracy of NO_(x) is increased by 3.7%and 2.6%respectively,and the prediction accuracy of CO_(2) emission is increased by about 3%.The prediction accuracy of BSFC is improved by 4%and 8%respectively.GP model predictes that the overall mean square relative error of NO_(x),CO_(2) and BSFC is less than 0.13,and the overall mean absolute percentage error is less than 0.1%.