Development of an Artificial Neural Network to Predict In-Use Engine Emissions

Abstract

A method to predict in-use diesel engine emissions is developed based on engine dynamometer and in-use data acquired at the West Virginia University Center for Alternative Fuels, Engines, and Emissions. (WVU CAFEE). The model accounts for the effects of road grade on generated emissions; a need for this model is evident in literature. Current modeling methods do not account for the effects of road grade, and have been shown to under-predict NOx by as much as 57%. It is determined through present research and a review of relevant literature that an artificial neural network (ANN) was the most applicable modeling method. A modular ANN was developed to predict the heavy duty diesel engine emissions. The two modules were trained independently, the first module was trained with data acquired through inuse testing, and the second module was trained with data acquired via engine dynamometer testing. The first module predicted the engine speed and torque associated with the inputs of road grade and vehicle speed, while the second ANN employed the first ANN's outputs, and predicts the emitted quantities of NOx, CO 2 , HC, and CO. A series of training and verification runs are conducted in order to determine the optimum ANN characteristics. Once the ANN was finalized, it was trained with and employed to predict the emissions associated with a variety of routes. When the ANN was trained with a combination of in-use and engine dynamometer data, the ANN is able to predict NOx emissions associated with that same route within 6% of the measured values. The average difference between the measured and predicted CO 2 values for the same training and verification scenario mentioned above was less than 15%. It was also demonstrated that the ANN was able to predict emissions that are associated with routes that differ from those by which it is trained. When the ANN was trained with in-use data from a specific route, it was able to predict the NOx and CO 2 emissions associated with a different route with percent differences from the measured values of 20% or less.