Research On Multi-perspective Attention Networks For Spatial-temporal Data Prediction

In recent years,with the continuous expansion of urban scale,the number of car ownership has been increasing rapidly.While cars bring convenience to people’s life,the problem of road congestion is becoming more and more prominent.Traffic flow is one of the important parameters to measure the road operation state.Accurately predicting the traffic flow in advance can effectively pre-allocate and schedule traffic resources,thus improving the road management level and maximizing the utilization rate of the road.It is of great significance to traffic management departments,drivers and passengers.Spatial-temporal data is highly complex,which is manifested in temporal dependence,spatial dependence,spatial-temporal dependence,time heterogeneity and local tendency.Most of the existing prediction models for spatial-temporal data can not fully consider the above five characteristics.In addition,the interference of noise in space and the influence of multi-source information need to be further explored.In order to improve the effectiveness of spatial-temporal data prediction,this paper studies the graph attention network model based on time heterogeneity and the graph attention network model based on multi-source information fusion respectively.The main research contents of this paper are as follows:First of all,most of the existing spatial-temporal data prediction models do not consider the different influences of absolute time and relative time interval on traffic flow,and often ignore the interference of noise in space.To solve the above problems,a timeheterogeneous attention mechanism is designed to model the influence of absolute time and relative time interval on traffic flow.In addition,considering the interference of noise,a dynamic noise filtering mechanism is designed.Finally,a Time Heterogeneous Graph Attention Networks(THGAN)model is proposed to deal with time heterogeneity and noise effectively.Secondly,the existing prediction models for spatial-temporal data often ignore the local time trend and seldom consider the feature aggregation of multi-source information.To solve these two problems,two strategies,local time trend modeling and multi-source information fusion,are proposed respectively,and a Multi-source Information Fusion Graph Attention Networks(MIFGAN)model is proposed.The model considers the trend effect of the nearest neighbor time slices,and can effectively integrate the influence of multi-source information(such as edge information and graph information)in the modeling process.Finally,extensive experiments are conducted on the public highway traffic flow dataset Pe MS.The experimental results show that the prediction performance of the THGAN model is better than the existing prediction methods,and the enhanced MIFGAN model further improve the prediction performance compared with the THGAN model.The experimental results verify the validity of the time heterogeneous attention mechanism and multi-source information fusion attention mechanism proposed in this paper,which can effectively mine the complex characteristics of spatial-temporal data and improve the prediction accuracy of spatial-temporal data.