LSTM-Based Predictive Control for Connected Autonomous Electric Vehicles Platoons

To successfully implement platooning of connected autonomous electric vehicles (CAEVs), it is necessary to actually cope with model uncertainties and unavoidable external disturbances. Indeed, in real-world driving scenarios, model mismatches and uncertainties due to variable conditions can severely impair the robustness of desired formation trajectory tracking, impacting the safety level, and also deteriorating the energy management. To tackle and solve this issue, we propose a framework where Long Short-Term Memory (LSTM) networks along with the Comprehensive Power-based Electric Vehicle Consumption Model (CPEM) are embedded into a Distributed Nonlinear Model Predictive Control (DNMPC) algorithm, in order to achieve the optimal ecological trajectory tracking. The accuracy of LSTM nets has been evaluated by leveraging the high-fidelity Mixed Traffic Simulator (MiTraS) co-simulation platform, where the performance of the CAEVs are analyzed in different realistic traffic scenarios and operative conditions. The performance of the proposed deep learning-based DNMPC controller, as well as its advantages w.r.t. model-based predictive approaches, are evaluated via numerical simulations. The results show that the proposed distributed control framework is effective in driving platoon formation towards the leader reference behaviour while ensuring an ecological and safe policy, even in the presence of modelling errors and unknown/unmodeled dynamics.