Abstract: We proposed a novel real-time vehicle speed prediction method based on Gaussian process regression (GPR) technology, which accurately and effectively predicted the velocity of the preceding vehicle while quantifying the uncertainty of the prediction. This method introduced a combination kernel function SEM of squared exponent (SE) and Matern, and improved the combination kernel function to SEM*. This effectively balanced the advantages and disadvantages of a single kernel function for vehicle speed prediction, and a particle swarm optimization method for real-time solution in hyperparameter optimization was adopted. The simulation analysis of 2 s vehicle speed prediction under transient operating conditions shows that under the FTP75 working  condition, compared to the radial basis SE kernel function with better single kernel performance, the SEM method reduces the mean absolute error (MAE) and root mean square error (RMSE) standards by 10.09% and 7.23% respectively, while the SEM* method reduces the two error indicators by 8.02% and 8.13% respectively compared to the SEM method. Under typical urban working conditions, the SEM reduces MAE and RMSE standards by 3.44% and 4.16% respectively compared to the SE method, while the SEM* reduces the two error indicators by 3.57% and 2.17% respectively compared to the SEM method. At the same time, the SEM* method reduces the maximum single calculation time relative to the SE method by 0.3 s under the FTP75 working condition, and the cost paid under typical urban conditions is an increase in the maximum single calculation time relative to the SE method by 0.015 s, but the calculation time is still within 0.1 s of the sampling time, which has real-time performance. 

Key words: combination kernel function, Gaussian process, vehicle speed prediction