Research On The Prediction Of Engine Performance And Emissions Using Artificial Neural Network

This article firstly introduced problems on energy consumption,relevant researches on alternative fuels,the application and advantages of ethanol as a fuel,the advantages and flexibility of combined injection engines,and the fundamental theories of neural networks.The way of fuel injection on the combined injection engine by direct injecting ethanol together with gasoline port injecting was determined,an engine measurement and control system matches the accuracy requirements was also built,where then an experimental procedure was carried out on.Data from the engine measurement and control system was used as the sources on the following researches.Python was used as the coding language for data processing and deep neural network model building.Plenty of parameter debugging,verification and model optimization work were performed.Totally 9 optimal models under their respective parameters were determined and trained.The model based on artificial neural network theory was realized on predicting specific values of performance and emission parameters of a combined injection engine fueled with ethanol and gasoline.Related analysis was then carried out.The 9 neural network models have excellent predictions on test data which have never experienced the training and verification process of neural networks without data leakage.Concretely,engine speed,throttle opening angle,intake pressure,intake temperature,ignition timing,port injection duration,direct injection timing,and direct injection duration were set as inputs,the average indicated pressure,effective power,torque,CO,NO_x,HC,and the number of particles were predicted by 7 neural network regression models with the number of parameters91265,91265,91585,91265,6882,68962 and 3201.On the test data,the correlation coefficient between predicted value and actual value respectively reached 0.983094,0.986104,0.985673,0.996829,0.981892,0.967135,0.973625.Mean absolute percentage error on test data respectively reached 0.016799%、0.026981%、0.022361%、0.332186%、0.151085%、3.161906%and 0.968660%.What’s more,the engine speed,throttle opening angle,intake temperature,intake pressure,ignition timing,intake port inject duration,direct injection time,and direct inject duration were used as inputs.Handling the data’s characteristics of the engine cylinder pressure,a model based on one-dimensional transposed convolutional neural network,one-dimensional convolutional neural network and up-sampling was established.The concrete numerical changes of the in-cylinder pressure in the entire cycle(in-cylinder pressure at every crankshaft angle)were predicted,with the parameters number of 62766415 and 237140 respectively.The average correlation coefficients on the test data are 0.997559 and 0.998685,mean absolute percentage errors are0.076363%and 0.088969%.Both models can predict the actual change of the engine cylinder pressure,and the characteristics that the prediction result can reflect more parameters than peak pressure,peak phase,rate of pressure rising,combustion start angle,etc.The predictions of engine performance and emission parameters based on neural network on various parameters are excellent.9 neural network models can be used to predict actual parameters within the allowable error range.At this point,the engine control parameters can be changed,and the actual values of the engine test parameters with higher accuracy can be predicted by inputting the control parameters to the trained neural network without any experimental procedure.This will save costs and the period on engine designing,researching,optimizing and analyzing.It will also assist or even replace the actual engine working process simulation based on physical and mathematical models,as well as other engine simulation methods except based on neural networks.