In modern transportation systems, the management of traffic safety has become increasingly critical as both the number and complexity of vehicles continue to rise. These systems frequently encounter multiple challenges. Consequently, the effective assessment and management of collision risks in various scenarios within transportation systems are paramount to ensuring traffic safety and enhancing road utilization efficiency. In this paper, we tackle the issue of intelligent traffic collision prediction and propose a vehicle collision risk prediction model based on vehicle-to-vehicle (V2V) communication and the graph attention network (GAT). Initially, the framework gathers vehicle trajectory, speed, acceleration, and relative position information via V2V communication technology to construct a graph representation of the traffic environment. Subsequently, the GAT model extracts interaction features between vehicles and optimizes the vehicle driving strategy through deep reinforcement learning (DRL), thereby augmenting the model’s decision-making capabilities. Experimental results demonstrate that the framework achieves over 80% collision recognition accuracy concerning true warning rate on both public and real-world datasets. The metrics for false detection are thoroughly analyzed, revealing the efficacy and robustness of the proposed framework. This method introduces a novel technological approach to collision prediction in intelligent transportation systems and holds significant implications for enhancing traffic safety and decision-making efficiency.