Emissions predictive modelling by investigating various neural network models

This paper presents a two-stage emissions predictive model developed by investigating common feedforward neural network models. The first stage model involves predicting engine parameters power and tractive forces and the predicted parameters are used as inputs to the second stage model to predict the vehicle emissions. The following gasses were predicted from the tailpipe emissions for a scooter application; CO, CO2, HC and O2. Three feedforward neural network models were investigated and compared in this study; backpropagation, optimization layer-by-layer and radial basis function networks. Based on the experimental setup, the neural network models were trained and tested to accurately predict the effect of the engine operating conditions on the emissions by varying the number of hidden nodes. The selected optimization layer-by-layer network proved to be the most accurate and reliable predictive tool with prediction errors of ±5%. The effect of the engine operating conditions on the tailpipe emissions for a scooter is shown to display similar qualitative and quantitative trends between the simulated and the experimental data. This study provides a better understanding in effects of engine process parameters on tailpipe emissions for the scooter as well as for general vehicular applications.