Abstract: In order to ensure the trajectory tracking accuracy and driving stability of autonomous vehicles, a model predictive control with variable predictive horizon method was proposed based on on-line identification of vehicle lateral stability state and fuzzy algorithm. Aiming at the online recognition of vehicle stable state, k-means clustering algorithm was used to cluster the parameters of vehicle driving state and obtain the cluster centroid. The real-time safety level of vehicle was obtained by comparing the Euclidean distance between the current vehicle state quantity and different cluster centroids online. At the same time, the lateral offset of the current vehicle track tracking is calculated. With the two as inputs, the variation of prediction time domain is calculated online by fuzzy algorithm and output to MPC controller to realize adaptive adjustment of prediction time domain. Finally, the optimal control sequence of the track tracking of the autonomous vehicle is solved to achieve the goal of high precision trajectory tracking control under the premise of maintaining vehicle stability. The results of CarSim/Simulink co-simulation show that the improved MPC algorithm is superior to the traditional MPC controller in improving the trajectory tracking accuracy and lateral stability of autonomous vehicles.