DHA-MPC: A Lightweight Dual-Horizon Adaptive Model Predictive Controller for Autonomous Vehicle Emergency Lane Changes under Variable Adhesion

Abstract In autonomous driving, traditional model predictive control (MPC) faces a fundamental trade-off between control accuracy and real-time computational efficiency. This limitation frequently results in hazardous conditions during high-speed maneuvers. To address the tension between tracking accuracy and computational efficiency in high-speed emergency evasive maneuvers, a dual-horizon adaptive model predictive control (DHA-MPC) framework adopting an offline-online collaborative architecture is proposed. Offline, an adaptive weighted particle swarm optimization (AW-PSO) algorithm globally optimizes the prediction and control horizons (Np, Nc). This optimization rigorously incorporates coupled constraints, including vehicle speed, variable road adhesion, and path curvature. Online, a lightweight multilayer perceptron (MLP) model serves as a sub-millisecond real-time inference engine to dynamically adjust the horizon parameters without updating complex prediction matrices. Through experimental verification under the extreme condition of a high-speed double lane change (DLC) maneuver at 70 km/h with a sudden variation in the adhesion coefficient, the DHA-MPC framework effectively limits the peak lateral error to 0.78 m. Furthermore, it significantly suppresses severe transient oscillations in the yaw rate and lateral acceleration. Hardware-in-the-loop (HIL) testing confirms the high determinism of the MLP forward inference time, recording an average execution time of only 26.8 seconds. Finally, Gaussian noise injection experiments further verify the trajectory tracking robustness and the active safety baseline of the system under the degradation of state observation.