Implementation of computationally efficient Taguchi robust design procedure for development of ANN fuel consumption prediction models

摘要

Reduction of passenger cars fuel consumption and associated emissions are two major goals of sustainable transport over the last years. Passenger car fuel consumption is directly related to a number of technological aspects of a given car, driver behaviour, road and weather conditions and, especially at urban level, road structure and traffic flow and conditions. In this paper, passenger car fuel consumption was assumed to be a function of three input variables, i.e. day of week, hour of day and city zone. Over the period of 6 months (during 2015) a car was driven in the randomly chosen routes in the city of Ni? (Serbia) in the period from 8 to 23 h. The fuel consumption data recorded through on-board diagnostics equipment were used for the development of Artificial Neural Network (ANN) models. In order to efficiently deal with a number of ANN design issues, to avoid usual trial and error procedure and develop robust, high performance ANN models, the Taguchi method was applied. For experimentation with ANN design parameters (transfer function, the number of neurons in the first hidden layer, the number of neurons in the second hidden layer, training algorithm), the standard L18 orthogonal array with two replications was selected. Statistical results indicate the dominant influence of the training algorithm, followed by the ANN topology, i.e. interaction of the number of neurons in hidden layers, on the ANN models performance. It has been observed that 3-8-8-1 ANN model represents an optimal model for prediction of passenger car fuel consumption. This model has logistic sigmoid transfer functions in hidden layers trained with scaled conjugate gradient algorithm. By using the Taguchi optimized ANN models, analysis of passenger car fuel consumption has been discussed based on traffic conditions, i.e. different days of the week and hours of the day, for each city zone and separately for summer and winter periods.