Research On Traffic Flow Forecasting Based On Multi-Graph And Multi-Factor Based On Spatial-Temporal Graph Networks

Traffic flow forecasting is one of the basic problems of intelligent transportation systems.Many applications are based on flow forecasting.However,due to the complex temporal and spatial correlation of traffic flow,it is very difficult to accurately predict traffic flow.With the rapid development of graph neural network,it is possible to accurately extract the spatial correlation of traffic flow.At present,there are two main types of spatial-temporal graph neural network model for traffic flow forecasting.One is to extract spatial and temporal correlations in a sequential manner,and model construction is realized by stacking modules,and the other is to first fuse spatial correlation and then extract temporal correlation for model building.For the first type,since the temporal and spatial correlations of different places at different times are different,and the temporal correlation and spatial correlation are closely intertwined,the potential interaction between them may not be extracted by modeling alone.For the second type,we believe that temporal correlation is more important for traffic forecasting.Extracting spatial correlation first will affect the extraction of temporal correlation.Aiming at the deficiencies of the spatial-temporal graph neural network model in traffic flow forecasting,firstly,this paper proposes a first-temporal-then-spatial fusion traffic flow forecasting model,which can dynamically extract the spatial-temporal correlation between different sensors at different times.Compared with the alone extract temporal or spatial model,the mean absolute error is reduced by 14%and 15%,respectively.Compared with the sequential spatial-temporal model,the mean absolute error is reduced by 9%.Compared with the first-spatial-then-temporal fusion model,the mean absolute error is reduced by 11%.Experiments prove that the spatial-temporal prediction model proposed in this paper has better prediction ability.Secondly,the construction of the spatial relationship graph is the key to extracting the spatial correlation.In this paper,starting from the two aspects of the local correlation of the road network and the similarity of the flow data,five sensor relationship graphs are constructed.It is found that the structure of the graph does affect the experimental results.A reasonable and accurate graph can improve the prediction results,and an unreasonable graph can even reduce the prediction results.The relative difference between the best and worst prediction results of different graphs reaches 8%.In addition,we also carry out the fusion of spatial relations and construct two fusion models,result fusion and graph fusion.Experiments show that the results of the fusion are better than the single graph prediction,and the result of result fusion is better than the results of graph fusion.Finally,traffic flow is affected by many factors.In addition to its own historical data and the spatial relationship of the road network,it will also be affected by holidays,weather,extreme events and traffic incidents.This paper discusses the impact of traffic flow parameters,time,and incidents on the flow forecasting,and finds that selecting appropriate extra factors can improve the prediction performance.For the multi-graph fusion model,after adding the time factors,the mean absolute error is reduced by 7%.