A trajectory privacy protection mechanism based on clustering and deep learning (PPCDL) was proposed aiming at the problem of privacy leakage faced by users in the trajectory distribution of Internet of Vehicles. The trajectory space was divided into multiple regions using timestamps by considering the time factor in the trajectory in order to obtain the distribution points of trajectories within each region. Improved stable membership multi-peak clustering was performed on each region, and the privacy budget matrix was pre-allocated based on the trajectory density of each region. The time graph convolutional network model was utilized to extract spatiotemporal features from trajectory data for training and predicting the pre-allocated privacy budget matrix. The trajectory data was perturbed by adding the appropriate Laplace noise based on the prediction results before it was published. The theoretical analysis and experimental results show that PPCDL has lower time overhead and can predict the privacy budget more accurately compared with the comparison mechanism. Laplace noise can be added to the trajectory data in a reasonable manner by using PPCDL, which effectively improves the availability of the trajectory data.