Future-heuristic differential graph transformer for traffic flow forecasting

Traffic Flow Forecasting (TFF) is crucial for various Intelligent Transportation System (ITS) applications, including route planning and emergency management. TFF is challenging due to the dynamic spatiotemporal patterns exhibited by traffic flow. However, existing TFF methods rely on the “average” spatiotemporal patterns for forecasting. To this end, this study investigates a heuristic-aware model named “Future-heuristic Differential Graph Transformer” (FDGT) for TFF with dynamic spatiotemporal patterns. Specifically, we define a kind of heuristic knowledge, called “future statistic” which provides reference information to describe the status of an object in the future. Then, we embed these statistics as coding features in the temporal domain of inputs. Next, we utilize Higher-order Differential Neural Networks (HDNNs) to enhance the perception of variation trends in the series. Moreover, we employ a Dual Spatiotemporal Convolutional Module (DSCM) to simultaneously learn global and local spatiotemporal dependencies. Finally, the Future-heuristic Fusion (FF) adaptively optimizes the weight distribution of each component, dynamically fuses the decoder's initial prediction and future statistics, and improves the model's generalization ability to capture spatiotemporal heterogeneities at different periods. Experimental results on four public datasets demonstrate that FDGT outperforms existing state-of-the-art TFF methods while maintaining superior execution efficiency.