A fast approximate model predictive control solving method for vehicle path tracking using discrete-time recurrent neural network

To address the response time-delay of the steering actuator and reduce the computational burden of model predictive control (MPC), this paper presents a fast approximate MPC scheme based on a discrete-time recurrent neural network (DRNN-MPC). First, the original convex quadratic programming (CQP) problem is reformulated into an approximate CQP with only boundary constraints, which is efficiently solved by the constructed DRNN, and the exponential stability of the DRNN is rigorously proven. An inverse model compensation strategy is then applied to compensate for the actuator’s response time-delay. Comparisons with state-of-the-art ACADOS and CVXGEN solvers show that the proposed DRNN-MPC significantly reduces calculation time while ensuring satisfactory tracking performance. The maximum MPC calculation time is reduced by 73.04%, and the average MPC calculation time is cut down by 85.96%, demonstrating excellent potential for engineering applications.