Intra-urban Region-based Traffic Flow Prediction Based On Spatial-temporal-semantic Graph Convolutional Network

Intra-urban region-based traffic flow prediction plays a significant role in traffic management and public safety.Therefore,in recent years,it has attracted wide attentions from many researchers.As a spatial-temporal prediction task,intra-urban region-based traffic flow prediction is challenge owing to the complex spatial and temporal dependencies in spatial-temporal data.The early traffic prediction studies that are based on the traditional statistical and machine learning models are difficult to model the spatial and temporal dependencies concurrently,which limits the prediction performance.Although some deep-learning-based traffic prediction methods can learn spatial-temporal features automatically and model the spatial and temporal dependencies concurrently,these studies cannot well utilize the spatially contextual information(e.g.,traffic flow changes of adjacent regions)and also overlook the semantic relationships among regions(e.g.,the correlation between regional traffic flow patterns).To address these challenges,this study proposed a novel deep learning model called spatial-temporal-semantic graph convolutional network(STS-GCN)for intraurban region-based traffic flow prediction.In the proposed STS-GCN,a sequential modeling strategy was utilized to model the spatial and temporal dependencies.The experiments with two public traffic flow datasets show that the proposed model STSGCN outperforms other baseline methods with RMSE reductions of 5.1% and 8.6% on the Bike NYC and Taxi BJ datasets,respectively,validating the effectiveness of the proposed prediction model.In addition,the ablative experiments further prove the effectiveness of the proposed STS-GCN.Our contributions are as follows:(1)When modeling the temporal dependency,most of the existing studies cannot well utilize the spatially contextual information.To address this challenge,this study proposed a spatial-context enhanced Long Short-Term Memory(SC-LSTM)to utilize the sequential and spatially contextual information in the temporal dependency modeling process.By utilizing the local spatial-contextual information of each region,the proposed SC-LSTM can improve the prediction performance.(2)When modeling the spatial dependency,most of the existing studies overlook the semantic relationships among regions.To address this challenge,this study designed a novel graph construction method to depict multiple relationships among various regions not only from an adjacent view but also a semantic view.The designed graph construction method helps ignore false signals from adjacent but uncorrelated regions and can model the long-range spatial dependency from a global spatial perspective,which help further improve the prediction performance.