A Deep Long Short-Term Memory Network Embedded Model Predictive Control Strategies for Car-Following Control of Connected Automated Vehicles in Mixed Traffic

This paper proposes a framework for deep Long Short-Term Memory (D-LSTM) network embedded model predictive control (MPC) for car-following control of connected automated vehicles (CAVs) in traffic mixed with human-driven vehicles (HDVs) and CAVs. The framework consists of: 1) lead HDV trajectory prediction through D-LSTM; and 2) CAV car-following control via MPC based on the predicted vehicle trajectory. For the trajectory prediction, two D-LSTM structures are developed based on the availability of preceding vehicle information: 1) ‘sufficient’ historical information of the position and speed of multiple vehicles ahead; and 2) ‘insufficient’ information where preceding vehicle information is unavailable (e.g., due to failed communication). Based on the prediction, a distributed MPC is designed for each scenario by incorporating the predicted trajectory into state space construction. The proposed D-LSTM models are trained and tested with the NGSIM data for validation. Numerical simulation results for various traffic conditions suggest that the proposed strategies perform better than traditional MPC methods in terms of control objective cost reduction, smoother control, and stabilizing effect. The results also indicate that the sufficient information case outperforms the insufficient information case as expected, which highlights the importance of stable communication.

I. Introduction

Connected automated vehicles (CAVs), equipped with advanced communication and automated functions, holds promise to improve microscopic vehicle maneuvers and consequently system-wide traffic performance. Particularly, the longitudinal control of CAVs brings a great opportunity to improve traffic flow throughput and stability (e.g., [1], [2], [3], [4], [5], [6], [7]) via precise vehicle control, cooperative sensing, and communication. To this end, many CAV car-following (CF) control algorithms, namely adaptive cruise control (ACC) or cooperative ACC (CACC; with communication function), have been developed aiming to improve the CF performance, fuel consumption, and stability (e.g. [3], [8], [9], [10]). Based on the control approach, the state-of-the-art control strategies can be largely categorized into three types: model predictive control (MPC) with explicit constraints (e.g., [11], [12], [13], [14], [15], [16]), closed form linear/nonlinear controllers (e.g., [17], [18], [19], [20], [21]), and data-driven approaches (e.g., [22], [23], [24], [25], [26]).