Research on Eco-driving Strategy in Urban Scenarios Based on Deep Reinforcement Learning

Abstract

As vehicle intelligence and connectivity develop by leaps and bounds, high attention has gradually been attracted to the eco-driving. It is significant to investigate eco-driving which is an effective method to improve both fuel economy and mobility in urban scenarios. However, most of the methods of current researches on eco-driving are model-based methods, in which complex traffic scenarios are a great challenge to the construction of models. Therefore, a deep reinforcement learning (DRL) algorithm which is model-free is applied to determine eco-driving strategy. Specifically, this paper proposes a method based on soft actor-critic (SAC) algorithm to realize eco-driving. Firstly, the problem of eco-driving is introduced according to the urban scene. On this basis, SAC algorithm is introduced and SAC model is designed to realize eco-driving. Finally, the method is verified by training and testing. The method based on SAC can effectively realize the eco-driving of vehicles according to the results. Compared with the global optimization algorithm such as simulated annealing (SA), energy consumption is reduced by 11.25%, while the whole decision-making time can be greatly reduced to 16.01%, which is of great significance for practical application.