Performance of neural network feature subset selection in injected fuel flow forecasting

This paper compares the performance of some feature subset selection in injected fuel flow forecasting. Injected fuel flow leads to an accurate measurement of car´s fuel consumption. The fuel consumption of a car depends on many factors like road, weather, and driver behaviour that are rigid for a car manufacturer to influence. Speed, stepper speed_step, stepper speed_angle, revolution, stepper rev_step, stepper rev_angle, fuel volume, stepper fuel_step, stepper fuel_angle, fuel transducer input, current fuel consumption, gear, distance to empty in volume, distance to empty in kilometre, current distance, and battery voltage are measured from the experimented car. The multilayered perceptron network trained by Levenberg-Marquardt algorithm was selected as a black box model for forecasting purposes. The input variables were taped from car sensors. The criterions for comparison are based on the coefficient of determination (R2). Three difference feature subset selection (A, B and C) consists of 1267 data samples have been collected. The first 700 data were used for training and the rest were used in validation and forecasting process. The three subset of the feature (A, B and C) are selected based on trial and error. The results show that feature subset selection B outperformed feature subset selection A and C significantly.