Abstract: High-precision vehicle speed prediction in cut-in scenarios is the key to ensuring the safety of autonomous driving cut-ins. To improve the safety of autonomous driving vehicles in cut-in scenarios, this paper studies the high-precision prediction method of ego-vehicle speed in cut-in scenarios based on vehicle-vehicle coupling risk clustering. First, the vehicle cut-in and cut-out segments are extracted based on the natural driving data obtained from the experiments, and the clustering analysis is performed based on the collision risks and acceleration correlation features using the K-means method. Second, based on the support vector machine (SVM) model, the online classification of vehicle-vehicle interaction state of cut-in and cut-out conditions is performed, and the real-time prediction of dangerous cut-in conditions is made. Finally, an improved vehicle speed prediction method based on ARIMA model (Autoregressive Integrated Moving Averaged Model) is proposed, optimizing real-time vehicle speed with online recognition results. Simulation results show the improved ARIMA vehicle speed prediction based on collision risk clustering significantly improves cut-in safety, cutting the vehicle collision risks by 10%～20% when compared to the traditional prediction methods. Our research may provide some references for improving the cut-in safety of autonomous driving vehicles.