Traffic forecasting remains a critical component in urban traffic management, especially as cities face unprecedented growth and increasing vehicle loads. Although graph-based models have demonstrated strong performance in traffic forecasting, they are not applicable when dealing with limited, non-graph-structured data. In this study, the goal is to do long term traffic prediction on target roads where only the traffic load of two adjacent roads and certain environmental-contextual factors are monitored. Therefore, the use of graph-based methods is unsuitable in this approach. This study proposes a novel context-aware CNN-GRU-Attention model specifically designed to overcome the limitations of graph-based methods for non-graph structured data. The model combines convolutional layers to capture local dependencies, GRU layers for temporal sequence learning, and a soft attention mechanism to emphasize long-term dependencies within the data. This architecture enables the model to effectively identify complex traffic patterns shaped by contextual factors such as weather conditions and holidays. Experimental evaluations on multiple datasets demonstrate that the proposed approach consistently outperforms traditional models, offering notable gains in predictive accuracy and robustness across various forecasting horizons. In summary, this work presents a scalable, adaptable solution to traffic forecasting in scenarios where graph structures are infeasible, advancing the field of urban traffic prediction through a context-sensitive methodology suited for modern traffic management applications. The source code for model implementing is publicly available.

Keywords