Platoon Communication Power Control Under V2V Data Uncertainty: A Robust DRL Approach

Connected and autonomous vehicle (CAV) platoon control has been considered as a promising technology to improve the safety and efficiency of intelligent transportation systems. In the platoon control decision-making process, vehicle-to-vehicle (V2V) communications are required to perform information exchange between vehicles. Existing studies commonly assume ideal V2V communication conditions, where the vehicles can receive completely accurate V2V data. However, in realistic scenarios, the V2V communication is subject to channel fading or communication attacks, which makes the vehicles to receive inaccurate or malicious V2V data, causing ‘V2V data uncertainty’. To tackle this problem, we investigate the CAV platoon control problem under the V2V data uncertainty. Firstly, we formulate a stability optimization problem under the V2V data uncertainty, based on an improved car-following model that incorporates multiple predecessors following-based V2V communication topology. Secondly, by employing linear stability theory in conjunction with the improved car-following model, we derive the stability constraint for the optimization problem. Then, by solving this optimization problem at every time step, an optimal control input is obtained to maintain system stability, thereby indirectly controlling the CAV platoon. To learn the optimal control strategy, we propose a robust deep reinforcement learning (DRL) approach. This approach demonstrates that the difference between the value function under optimal state perturbation and that under unperturbed conditions is bounded by a tunable upper limit. Since data uncertainty is intuitively reflected in the state, adjusting this upper bound enables the robust DRL approach to counteract the effects of V2V data uncertainty. Finally, simulation results indicate that the proposed approach significantly outperforms baseline approaches and the average allocation method in terms of platoon state control, stability, and comfort. Moreover, it is markedly superior to the baseline approaches regarding convergence and overall model performance.