Machine Learning Enabled Traffic Prediction for Speed Optimization of Connected and Autonomous Electric Vehicles

Abstract

Connected and autonomous vehicles (CAVs) can bring in energy, mobility, and safety benefits to transportation. The optimal control strategies of CAVs are usually determined for a look-ahead horizon using previewed traffic information. This requires the development of an effective future traffic prediction algorithm and its integration to the CAV control framework. However, it is challenging for short-term traffic prediction using information from connectivity, especially for mixed traffic scenarios. In this work, a novel machine learning enabled traffic prediction method is developed and integrated with a speed optimization algorithm for connected and autonomous electric vehicles. The traffic prediction is based on a hybrid macroscopic traffic flow model, in which the most challenging nonlinear terms are modeled with neural networks (NNs). The traffic prediction method can be readily applied to various mixed traffic scenarios. Information from connected vehicles is used as partial measurement of the traffic states and the rest unknown traffic states are estimated using a state observer. Then, the preceding vehicle's future trajectory is obtained to formulate the car-following distance constraint of the energy optimization problem. In a simulated scenario of 70% penetration rate of connectivity, the NN-based traffic prediction algorithm can reduce the root-mean-square errors of the prediction of preceding vehicle speed by near 50%, compared to the conventional traffic flow model. The energy benefit is 12.5%, which is satisfactory compared to 16.5% of the scenario with perfect prediction.