Spatial-Temporal Graph Neural Network Framework with Multi-source Local and Global Information Fusion for Traffic Flow Forecasting

As the increasing population is causing a large amount of traffic congestion nowadays, accurate traffic flow forecasting (TFF) has become increasingly significant for smart cities. While some recurrent neural network-based models have taken the spatial-temporal (ST) features of traffic flow data into account for TFF, they still treat the ST features from multi-source time-series data of each timestep separately, which may be inefficient to learn a general and robust representation with spatial correlation information for TFF. To address these issues, this paper proposes a Spatial-Temporal Graph Neural Network Framework (STGNNF) for TFF. This framework is based on the assumption that when a small range of adjacent temporal features of the multi-source time-series data is fused as the input, the model is still able to effectively learn meaningful representation and capture the time-series information. By doing so, the learned representation from the fused adjacent data can be more general and robust for TFF, because it considers the multi-source time-series data within a larger time scale. Moreover, three novel designs are proposed and integrated to further enhance the STGNNF, which are 1) a local ST unit for learning the local ST information from the fused adjacent multi-source time-series data; 2) a relevance evaluation module for paying more attention to the significant local ST information; and 3) a global ST unit for generating the global ST representation and information from local ST information with corresponding attentions. Experimental studies are conducted in three real-world traffic flow datasets, which indicate that STGNNF performs better than existing approaches and predicts reasonably when encountering anomalous data.