Research On Support Vector Machine Regression Model Based On Steady Condition Calibrationof A HCNG Engine

Hydrogen enriched compressed natural gas(HCNG)can improve engine efficiency and reduce engine emissions compared to conventional fuels,which is regarded as a feasible clean alternative fuelduring the transition from traditional vehicles to electric vehicles.As the engine electronic control system develops,the work of engine calibration has become more and more complex.WP6NG240E5 natural gas engine is involved as a base engine in this paper.With 20%(volume percentage)HCNG fuel,equivalent fuel-air ratio and ignition timing under steady condition of the engine are calibrated,covering full conditions of the engine.Support vector machine regression(SVMR)then is use to establish the mathematical model between engine performance parameters and operation parameters based on these steady condition calibration data,aiming at the guidance of the work of engine calibration.According to the steady condition calibration data,fuel-air ratio and ignition timing characteristics of HCNG engine are analyzed.Through the comparison in fuel-air ratio,ignition timing and various engine performance parameters between the 20%HCNG and the original CNG engine under wide open throttle condition,the effects and mechanism of achieving lean burn,improving engine efficiency and reducing NOx emissions by blending hydrogen are analyzed.Based on all the steady condition experimental data,engine speed,intake manifold pressure(MAP),equivalent fuel-air ratio and ignition timing are selected as independent variables,and output torque,equivalent CNG consumption(BSFC)and NOx emission(BSNOx)are respectively chosen as dependent variables.SVMR method is used then to establish the mathematic model between the independent and dependent variables.Grid search,particle swarm optimization(PSO)and genetic algorithm(GA)are respectively adopted to get the best models.Prediction results of the best models show that the mean absolute percentage errors(MAPE)of torque and BSFC prediction are within 3%,with the maximum relative error of about 6% for torque and 15% for BSFC,while the MAPE of BSNOx prediction,however,is over 10%,with the maximum relative error of as much as 60% or so,which is unsatisfactory.Research on the effects on prediction accuracy of the SVMR model for BSNOx by model parameters and training sample are conducted.The best SVMR model with high precision for BSNOx under small sample is obtained.Under the condition that the training sample is reduced to 1/3 of the full training sample,MAPE for BSNOx prediction of the model is decreased to 8.3%,with the maximum relative error of 25.9%,which is improved obviously.To comparethe prediction performance of SVMR with neural network(NN)for BSNOx,3 layers BP neural network is used for regression analysis based on the full data sample and the sample for best SVMR model.It is concluded that SVMR is better than 3 layers BP neural network on prediction performance,sensitive to the training sample,availability to the best model and stability of the best model,especially showing great potential on nonlinear regression problems with small sample.