Recurrent Neural Network Based Driving Cycle Development for Light Duty Vehicles in Beijing

Abstract

Abstract This paper presents a data driven, Recurrent Neural Network (RNN) based technique that models real world driving patterns without domain knowledge to develop driving cycles for light duty vehicles (LDVs) in Beijing. In contrast to approaches that feature sub-cycle selection from the original data set through Markov process and data clustering, our method models the conditional probability distribution of vehicle speed with RNN, in which the driving cycles are generated step by step. As a consequence, the presented method excludes the necessity for domain knowledge based feature extraction and corresponding data vectorization during modelling stage. In the end of this paper, the proposed method is evaluated through comparisons between synthesized driving cycle and the original data set based on 14 metrics. As the final results suggest, both of the cycles obtained from the two models established are of relatively high power demand compared with the original data set, while one of the model is able to yield driving cycles that are more inclusive in terms of velocity levels with less candidates, which, to some extent, makes it more suitable for emission tests compared with the other one.