Research On Spatio-Temporal Multi-Scale Convolutional Network For Traffic Forecasting

The study of traffic flow forecast is of great significance to the realization of resource allocation,energy utilization,traffic control and intelligent transportation system,and its essence is a kind of spatio-temporal forecast.Deep learning is applied in various fields and also plays an important role in the task of traffic prediction.When simulating complex urban-scale population flows,it is necessary to analyze the temporal and spatial correlations of historical data.At present,Recurrent Neural Networks（RNN）are widely used in temporal correlation modeling,and Convolutional Neural Networks（CNN）are widely used in spatial correlation modeling.However,these excellent deep learning models model the temporal and spatial characteristics of traffic data by superimposing or combining complex neural network structures,resulting in complex models and high computational costs.In addition,traffic transfer events may occur in any two areas of the city,forming a large-scale spatial correlation in a short time interval.Existing methods have not achieved satisfactory results when dealing with these two problems.Therefore,the research on traffic flow prediction based on temporal and spatial correlation needs to be further improved.This paper proposes a new spatio-temporal multi-scale convolutional network（STMSCN）to solve the problem of traffic flow prediction,and focuses on the analysis of the temporal and spatial correlation of traffic flow.First,in order to directly capture the spatial correlation and multi-scale characteristics of urban traffic flow,a new MSC unit is proposed,which can directly capture the spatial correlation and multi-scale characteristics of the entire city.The MSC unit eliminates the need for complicated deep network structure for spatio-temporal characteristic learning.In addition,so as to decrease the parameters of the model,connect the relationship between different timelevel features,and obtain the degree of influence between the features,we conducted an early fusion of the three temporal correlations（i.e.,recent segment,period segment and trend segment）of traffic data.Further,we will merge the low-level features learned through the convolutional layer with the high-level features learned by the MSC,and add the influence of external features at the end of the model.Finally,We avoid model overfitting.In this way,our proposed architecture can not only better model spatiotemporal features,achieve better prediction results and generalization capabilities,but also perform better in time consumption than existing models.The contributions of this article are summarized as follows:（1）We designed a new MSC unit,which can directly capture the spatial correlation and multi-scale characteristics of the entire city’s traffic flow data.（2）We propose a concise ST-MSCN architecture and early fusion mechanism,which can effectively reduce the scale of the model and training parameters.（3）We propose a new multi-level feature fusion strategy,which can combine lowlevel surface features and high-level abstract features at the same time to make the prediction of peak points and abnormal points more accurate.（4）We evaluated the method using traffic flow data in Beijing and New York.STMSCN is significantly better than the other nine methods in predicting traffic flow,and is far superior to the latest technology in terms of time consumption.