Translation of Distance-Specific Emissions Rates between Different Heavy Duty Vehicle Chassis Test Schedules 2002-01-1754

When preparing inventory models, it is desirable to obtain representative distance-specific emissions factors that truthfully represent the vehicle activity on a particular road (facility) type. Unfortunately, emissions values are often measured using only one test schedule, which represents a single average speed and a specific type of activity. This paper investigated the accuracy of predicting the emissions for a test schedule based on measurements from a different test schedule for the case of a medium heavy-duty truck. First, the traditional Speed Correction Factor (SCF) approach was examined, followed by the use of a power-based model derived from continuous data, followed by an artificial neural network (ANN) approach. The SCF modeling used distance-averaged emissions and cycle-averaged vehicle speed to predict distance-averaged NOx. The power-based modeling was based on linear and polynomial correlations between continuous axle power and NOx. The ANN models were trained on axle speed, torque, and their respective derivatives to predict continuous NOx. Artificial neural networks can identify highly non-linear relations between multiple input and output data, making them well suited for the task of emissions prediction. Previous research at West Virginia University used ANN's (trained on engine emissions data) to predict continuous heavy-duty chassis emissions. In the present study, data were obtained from a 1995 GMC box truck exercised through 16 different transient chassis test schedules. The truck was equipped with an automatic transmission and powered by a 127 kW (170 hp), 6.6 liter, 3116 Caterpillar engine with mechanical fuel injection. Simulated test weight for the truck was 9,980 kg (22,000 lb). Emissions characterization was performed using a heavy-duty chassis dynamometer, a full-scale dilution tunnel, continuous analyzers for gaseous emissions, and filters for PM emissions. The SCF approach yielded an average error of 26.7%, the power-based approach yielded 19.1% error, and the ANN approach yielded an average 10% error. It was also observed that some test schedules are more easily predicted than others, and some are a better source of training than others. The success of this modeling lends hope to the use of continuous on-board emissions measurement data to project the emissions for some target vehicle operation that cannot be achieved precisely while driving on the road.