Spatio-temporal Features Analysis And Its Application Based On Graph Convolutional Networks

A network topology is a common spatial structure composed of multiple nodes,where the attributes of each node not only vary over time,but also are influenced by the spatial correlation between nodes.The spatio-temporal data generated by network topology,as a new type of data,contains valuable information of time and geographic location.The trend of things,the chronological relationship and the evolutionary pattern can be understood by analyzing the temporal information.And the spatial distribution and intrinsic spatial relationship can be obtained by analyzing the geographic location information.The spatio-temporal patterns and features provide accurate support for decision-making.In the era of Big Data,spatiotemporal features have become one of the hot spots for research in various industries,especially in the fields of urban planning,transportation management,environmental monitoring,weather forecasting,earthquake early warning,etc.However,there are challenges in the analysis of spatio-temporal features that can lead to less-than-ideal results of prediction.First,there is still great challenge in effectively constructing spatial topology networks in non-Euclidean spaces.Based on the assumption of spatial proximity,most methods use grid partitioning and convolutional neural networks to construct spatial topology networks and extract spatial-temporal correlations.But this assumption often deviates from reality.Therefore,constructing and mining its intrinsic spatial correlation is not negligible.Secondly,dynamic spatial correlations bring difficulties to capture spatio-temporal features precisely.It is unreasonable to rely on a static or predefined spatial network to obtain spatial-temporal correlation for continuously changing time-series features.Therefore,to adapt the spatial network to the changing temporal features,it is necessary to dynamically adjust the spatial correlations according to the fluctuating temporal features to improve the accuracy of spatial-temporal feature mining.To address the above two challenges,two spatio-temporal feature analysis models are proposed based on deep learning algorithms for the traffic flow prediction problem,achieving effective improvement in prediction accuracy.The main research contents of this paper are as follows.(1)Spatio-temporal feature analysis model is generated by static spatial networks.In order to address the challenge of effectively constructing a spatial topology network in non-Euclidean space,multi-graph spatio- temporal graph convolutional network(MSTGCN)is developed for the transportation field and features fast training and prediction capabilities. The model combines graph convolutional networks(GCN)and temporal convolutional networks(TCN)to construct spatial network based on three different semantics to capture spatio-temporal correlations and improve accuracy.The model’s predictive accuracy is validated in an actual transportation system.By comparing with other baseline models, MSTGCN shows a significant improvement in traffic flow prediction accuracy.(2)Spatio-temporal feature analysis model is generated by dynamic spatial networks.To address the difficulty of accurately capturing dynamic spatio- temporal features,a spatio-temporal heritable neural networks(STHNN) model is proposed that can dynamically capture correlations in spatio- temporal networks with high prediction accuracy.However,the training and prediction processes are time-consuming,making it more suitable for prediction systems with high accuracy requirements but lower real-time requirements.STHNN introduces a new spatial hidden state within Gate Recurrent Unit(GRU)to record spatial inheritability,which achieves adaptive generation of spatial structure at different moments,according to the temporal feature states.By integrating GRU and GCN,STHNN model can obtain spatio-temporal correlations at any time step,and its prediction accuracy is verified on the Pe MSD4 and Pe MSD8 public datasets.(3)A prototype system is designed to implement a traffic flow prediction module,where prediction results are displayed by visualization tools.The accuracy of the model’s exploration of spatio-temporal features is analyzed through a comparison of predicted values with ground truth.Based on historical spatio-temporal features,the system predicts future traffic flow for all stations,displays and labels hotspots on a map,and provides effective data support for road managers.The two models,named MSTGCN and STHNN respectively,are proposed.They mainly address the difficulties of constructing spatial topological networks effectively in non-Euclidean space and capturing dynamic spatial-temporal relationships.Experiments and cases show that the above two models perform well in the spatio-temporal feature analysis task and can obtain accurate prediction results.In the prototype system,the traffic flow prediction of highway toll stations is achieved by the above two models,which provides effective data support for road managers.