Abstract:

A predictive Cooperative Adaptive Cruise Control (CACC) method was proposed based on the Physics-Informed Neural Network (PINN) and a longitudinal behavior prediction algorithm for human-driven vehicles (HDV) was developed, to address the behavioral uncertainty of HDV in mixed traffic environments including intelligent-connected vehicles. The PINN-based HDV behavior prediction model was constructed, in which an optimization problem with differential-equation constraints was transformed into a neural network parameter fitting problem. The predicted states of HDV were used as reference inputs to design a predictive CACC controller for mixed traffic based on Model Predictive Control (MPC) with soft constraints, referred to as PINN-MPC. The proposed controller was validated through simulations on the HighD dataset. The results show that the PINN controller with physical information improves the acceleration prediction accuracy by 28.9% within a 3 s prediction horizon compared with neural networks without physical information. Therefore, the proposed cruise control strategy enhances both driving safety and ride comfort.

Key words: mixed traffic, intelligent-connected vehicles, human-driven vehicles (HDV), cooperative adaptive cruise control (CACC), physics-informed neural network (PINN)