Taxi Traffic Flow Prediction Based on Spatiotemporal-Fusion Graph Neural Networks

Accurate short-term traffic flow prediction in complex urban road networks is of great significance for capacity organization and dispatch optimization in intelligent transportation systems. Using publicly available historical taxi trip records released by the New York City Taxi and Limousine Commission from January to June 2016, this study develops a spatiotemporal fusion framework for short-term traffic flow prediction. To address the nonlinearity, sparsity, and complex spatiotemporal dependencies of traffic flow sequences, the raw trajectory data are first cleaned, spatially gridded, and temporally discretized. Based on the spatial adjacency relationships among grid nodes, a graph structure is then constructed, and a serially coupled graph convolutional network and long short-term memory model is developed to capture spatial dependency features and temporal dynamic features, respectively. Experimental results on the New York City taxi dataset show that, compared with baseline models including the historical average model, long short-term memory network, graph convolutional network, and Transformer, the proposed model achieves better performance in terms of mean absolute error, root mean square error, and coefficient of determination. Furthermore, the SHAP (SHapley Additive exPlanations) method is employed to ANALYZE the differences in feature contributions across nodes in different functional zones from both temporal and spatial perspectives. The results indicate that the model exhibits heterogeneous temporal dependency depths and spatial aggregation patterns across different types of regions within the study area. In addition, regions with high feature contributions show a certain degree of spatial correspondence with the major traffic corridors in Manhattan, suggesting that the model is able to capture part of the spatiotemporal correlation structure of traffic flow in this dataset. Finally, the limitations of the proposed method in terms of static graph structure, response to extreme events, and integration of external factors are discussed. It should be noted that these findings are derived from New York City taxi data from the first half of 2016, and their generalizability to other cities, time periods, or traffic scenarios remains to be further validated.