Anticipatory Longitudinal Vehicle Control using a LSTM Prediction Model

In this paper, an approach for longitudinal vehicle control is proposed that integrates a Machine Learning model for the speed prediction of the leading vehicle in order to improve the energy efficiency of the control in a city driving environment. The prediction is employed in an additional control mode that extends a conventional state-of-the-art speed and headway control. The approach aims to reduce the vehicle consumption through a more anticipatory driving style. The prediction model uses an encoder-decoder LSTM network with additional information from Vehicle-2-X communication and is trained through supervised learning with training data generated from simulation. A co-simulation environment comprising of an ego vehicle simulation and a microscopic traffic simulation is used for the generation of training data and the assessment of the control approach. The proposed control is compared to a benchmark Adaptive Cruise Control scheme for three battery electric vehicles with different powertrain specifications on multiple routes in a simulated city-driving environment of Darmstadt, Germany. The results show that the consumption of the vehicles can be reduced for all vehicles by 3–5 % while still maintaining similar mean speeds on different routes through the city.