Objective: Accurate identification of road sections that critically impact the efficiency of urban road network operation is an important prerequisite for optimizing traffic resource allocation and improving road network resilience. Traditional identification methods often rely on topological attributes and ignore the time-varying characteristics of traffic states, which makes it difficult to capture key sections across different periods. Therefore, this study proposes a key link identification model that integrates structural and feature information, aiming to provide a more accurate basis for decision-making in urban traffic control and emergency management. Methods: In this study, a key road segment recognition model based on a graph attention network (GAT-KSI) is constructed. First, given the time-varying nature of traffic, GAT is used to obtain a feature influence score that quantifies the influence intensity of the running state between neighboring road sections. Second, by combining topological information from neighboring road sections, a structural influence score quantifying the potential for traffic flow to propagate between them is obtained. Then, the feature influence score and the structural influence score are combined to measure the feature-structure fusion influence between neighboring road sections in the road network. On this basis, the PageRank model is improved by using a fusion influence score between neighboring road sections to rank key road sections. Finally, based on road network data from Shenzhen, China, a load-capacity model of link cascading failure is constructed, and the congestion failure process of key links under two typical traffic situations, peak and flat peak, is simulated. The network efficiency index is used to evaluate the impact of congestion failures in key sections identified by different methods on the operation efficiency of the road network, thereby reflecting the ability of each method to identify these sections. Results: The experimental results show that the first 20 key sections identified by the GAT-KSI model can cause the road network efficiency to decrease by 55.9% during the peak period and 50.0% during the flat peak period when congestion failures occur. By contrast, the first 20 critical links identified by baseline methods such as degree ranking, betweenness centrality ranking, the traditional PageRank model, the gravity model, graph convolutional network, graph sample and aggregate, and GAT v2 only resulted in a decrease in road network efficiency of 20.0% to 43.9%. This result verifies that the GAT-KSI model can more accurately identify key road sections that have a significant impact on the operation of the road network under different traffic conditions, demonstrating stronger scene adaptability and recognition stability. Conclusions: The GAT-KSI model proposed in this study achieves accurate identification of key road sections by integrating road network topology and traffic time-varying characteristics. The model not only considers the dynamic propagation characteristics of traffic states but also accounts for the potential impact mechanisms of the road network structure, and it shows superior recognition performance across different traffic scenarios. The model can provide effective technical support for key node identification, congestion prevention, and emergency response in urban traffic management, offering important theoretical value and practical benefits.