Motor vehicle emission prediction models have been shown to underestimate emissions by factors of 2 to 3 in some cases. There are many reasons why current models do not predict automobile emissions under ‘real-world’ driving conditions accurately. Among them are: non-representative driving activity used to derive emission prediction algorithms; lack of driving activity input variables; statistical shortcomings of models; and non-representativeness of tested vehicles compared to on-road vehicles.
Given the inaccuracy of current emission prediction models and the need to accurately assess transportation control measures, incremental transportation supply changes, and intelligent transportation technologies, new activity-based emission prediction algorithms are required. Improved algorithms need to be sensitive enough to capture the effects of microscopic flow adjustments, or flow smoothing, that are now commonly considered among transportation and air quality planners.
This paper first presents some results of second-by-second emissions data collected for a 1991 Toyota Camry in Australia. This data is used to demonstrate the importance of modal activity, and the wisdom of incorporating modal variables into an emissions model. The data was used to develop an ‘elemental’ model for prediction of emissions of CO, HC and NOx; however, only the models for acceleration and deceleration CO emissions are presented here.
As a way to encapsulate the modal modeling approach into the existing modeling framework, a modal model estimated using traditional ‘bag data’ is introduced. The new model, comprised two linear regression modal models, collectively dubbed DITSEM (Davis Institute of Transportation Studies Emission Model), employs modal explanatory variables such as acceleration, positive kinetic energy and proportion of cycle at idle to predict CO emissions from both ‘high’ and ‘normal’ emitting vehicles.
To measure the model performance of DITSEM, CO emission prediction algorithms embedded in ‘competing’ models (CALINE4 and EMFAC7F) are used to predict emission test results over a wide range of driving cycles. Measures of model performance compared are mean squared error, mean absolute error, Theil’s U-Statistic and the linear correlation coefficient. Statistical comparisons show that DITSEM CO prediction algorithms are superior and capture the effect of microscopic flow changes well.
The authors suggest that significant interim improvements can be made using the existing ‘bag’ collected data⎛making them sensitive to microscopic changes in travel behavior. A more sophisticated modeling approach, one based on second by second data and similar to the one presented here, could be used for a more long term model improvement program. This large scale second-by-second effort should only be undertaken with specific measurement and modeling objectives in mind, since a great deal of improvement can already be achieved with the current data.