Sustainable DDPG-Based Path Tracking for Connected Autonomous Electric Vehicles in Extra-Urban Scenarios

This paper addresses the path-tracking control problem for Connected Autonomous Electric Vehicles (CAEVs) moving in a smart Cooperative Connected Automated Mobility (CCAM) environment, where a smart infrastructure suggests the reference behaviour to achieve. To solve this problem, a novel energy-oriented Deep Deterministic Policy Gradient (DDPG) control strategy, able to guarantee the optimal tracking of the suggested path while minimizing the CAEVs energy consumption, is proposed. To this aim, the power autonomy, the battery state of charge (SOC), the overall power train model -comprehensive of the electric motor equations, inverter dynamics and the battery pack model- is embedded within the training process of the DDPG agent, hence letting the CAEV to travel according to the best sustainable driving policy. The training procedure and the validation phase of the proposed control method is performed via an own-made advanced simulation platform which, combining <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Matlab &amp; Simulink</i> with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Python</i> environment, allows the virtualization of real driving scenarios. Specifically, the training process confirms the capability of DDPG agent in learning the safe eco-driving policy, while, the numerical validation, tailored for the realistic extra-urban scenario located in Naples, Italy, discloses the capability of the DDPG-based eco-driving controller in solving the appraised CCAM control problem despite presence of external disturbances. Finally, a robustness analysis of the proposed strategy in ensuring the ecological path tracking control problem for different CAEV models and driving path scenarios, along with a comparison analysis with respect model-based controls, is provided to better highlights the benefits/advantages of the proposed Deep Reinforcement Learning (DRL) control.