Neural Network Modeling of Emissions from Medium-Duty Vehicles Operating on Fisher-Tropsch Synthetic Fuel 2007-01-1080

West Virginia University has conducted research to characterize the emissions from medium-duty vehicles operating on Fischer-Tropsch synthetic gas-to-liquid compression ignition fuel. The West Virginia University Transportable Heavy Vehicle Emissions Testing Laboratory was used to collect data for gaseous emissions (carbon dioxide, carbon monoxide, oxides of nitrogen, and total hydrocarbon) while the vehicles were exercised through a representative driving schedule, the New York City Bus Cycle (NYCB). Artificial neural networks were used to model emissions to enhance the capabilities of computer-based vehicle operation simulators. This modeling process is presented in this paper. Vehicle velocity, acceleration, torque at rear axel, and exhaust temperature were used as inputs to the neural networks. For each of the four gaseous emissions considered, one set of training data and one set of validating data were used, both based on the New York City Bus Cycle. Four different types of artificial neural networks were investigated: linear, single hidden layer with sigmoid activation function, nonlinear polynomial (Sigma Pi), and Gaussian radial basis function neural network. The accuracy of the continuous estimates of the models was evaluated in terms of integral, maximum, and standard deviation of the modeling error. The accuracy of the integrated estimate over the entire cycle was also analyzed based on the percentage error. The model based on the radial basis function neural network provided overall better accuracy; however, the others remain viable alternatives due to their simpler structure and because their internal parameters can potentially reveal additional information on the mechanisms relating emissions and vehicle operation regime.