A deep learning lane-changing decision framework with wide spatiotemporal conditions for connected and automated vehicles

Currently, research rarely considers the impact of communication, especially for the large-scale communication environment, in the change-lane decision of connected and automated vehicles (CAVs) in mixed traffic composed of CAVs and human-driven vehicles (HVs). Thus, we build a Wide-Spatiotemporal Lane-Changing Decision Framework (WST-LCDF) with learning-based models for CAVs to increase the overall mixed traffic throughput, i.e., the average velocity of all vehicles. This framework considers a wide spatiotemporal relation between CAVs, so CAVs can obtain the surrounding physical information and communicate with other CAVs in cyber flow. After using the simulation data to train the framework, we find that the prediction accuracy of average velocity increases exponentially with the CAV market penetration rate (MPR). The prediction accuracy of WST-LCDF is higher than the Convolutional Long Short Term Memory Neural Network (Conv-LSTM) \ Full Connected Long Short Term Memory Neural Network (FC-LSTM). Use trained WST-LCDF to make lane-change decisions for CAVs in the simulation, and the results indicate that our network structure is better than the Convolutional Long Short Term Memory Neural Network (Conv-LSTM) \ Full Connected Long Short Term Memory Neural Network (FC-LSTM) in handling spatiotemporal correlations. We find that using WST-LCDF increases the average velocity by about 7.1%. In more than 85% of lane-change cases, WST-LCDF can help CAVs make lane-changing decisions to a higher overall velocity.