详情

原文传递 Predictability of Vehicle Fuel Consumption Using LSTM: Findings from Field Experiments

题名：	Predictability of Vehicle Fuel Consumption Using LSTM: Findings from Field Experiments
正文语种：	eng
作者：	Guanqun Wang;Licheng Zhang;Zhigang Xu;Runmin Wang;Syeda Mahwish Hina;Tao Wei;Xiaobo Qu;Ran Yang
作者单位：	School of Information Engineering Chang'an Univ. Xi'an 710064 China;School of Information Engineering Chang'an Univ. Xi'an 710064 China;School of Information Engineering Chang'an Univ. Xi'an 710064 China;School of Information Engineering Chang'an Univ. Xi'an 710064 China;School of Information Engineering Chang'an Univ. Xi'an 710064 China;Shaanxi Lingyun Electronics Group Co. Ltd. No. 1 Yuquan Rd. Baoji 721006 China;Dept. of Architecture and Civil Engineering Chalmers Univ. of Technology Gothenburg SE-412 96 Sweden;School of Information Engineering Chang'an Univ. Xi'an 710064 China
关键词：	Fuel consumption prediction; Long short-term memory (LSTM) neural networks; FuelNet; Eco-driving; Test site data; Real-world data
摘要：	It has been well-recognized that driving behaviors significantly impact the fuel consumption of vehicles. To explore how well deep learning methods can predict fuel consumption precisely and efficiently and then guide drivers to go in an energy-saving way, we propose a fuel consumption prediction model, namely FuelNet, based on long short-term memory (LSTM) neural networks in this study. First, we develop the proposed FuelNet model with numerous vehicle kinematics data and corresponding fuel consumption data collected in the test field and real-world scenarios. And we analyze the relationship between the prediction accuracy and different combinations of input variables, training set size, and the sampling interval of the raw data. Second, we conduct intensive field tests to demonstrate the applicability of our model to fuel consumption prediction for different speed conditions and vehicle types. Furthermore, the superior prediction performance of FuelNet is shown by comparing it with five other types of models, such as the physical model, statistical and regression model, conventional neural networks model, and other deep learning models. Finally, we apply it to three real case studies, which verify that FuelNet can precisely predict fuel consumption for different driving trajectories in many scenarios such as signalized intersection (average value of RE is 0.049), campus environments (RE is 0.030), urban roads (RE is 0.077), and highways (RE is 0.097), as well as can contribute to detecting abnormal fuel consumption.
出版年：	2023
期刊名称：	Journal of Transportation Engineering
卷：	149
期：	5
页码：	04023030.1-04023030.16