In order to realize the active prevention and control of rear-end collision on sharp curves, a dynamic prediction method of rear-end conflicts was proposed based on converged multi-source data. In this study, firstly, a model for distinguishing rear-end conflicts on sharp curves and a classification criteria of conflict level threshold were proposed by using the vehicle traveling data collected by drones and millimeter wave radar. The spatial distribution characteristics of such conflicts on sharp curves was subsequently analyzed. Then, 13 variables related to traffic flow characteristic such as vehicle type, ratio of large vehicle, and speed difference between sections were selected as input variables, and dynamic combined prediction models of rear-end conflicts on sharp curves with BP neural network, random forest and support vector machine algorithms were constructed based on particle swarm algorithm, respectively. The prediction performance of each prediction model was evaluated based on confusion matrix and area under curve, and the black box interpretation method was used to analyze the significant influence variables and the probabilities of rear-end conflicts occurrence. The results show that the TTC values of rear-end conflicts on sharp curves are smaller than those on flat or normal curved sections, and such conflicts are more serious on the gentle curve exit sections, and the conflicts risk on the inner side of the curves is the highest; the particle swarm algorithm-random forest model has the best performance in predicting the rear-end conflicts, with a sensitivity of 90.70%; the impact of average vehicle headway on rear-end conflict in sharp curve sections is the most significant. When the average vehicle headway is around 25 meters, the probability of conflict is the lowest. Factors such as mean centripetal acceleration and mean velocity also have a significant influence on it.