Traffic Demand Prediction Based On Spatio-Temporal Dynamic Multi-Graph

The development of the city leads to the increasing pressure of traffic operation.It is urgent to solve various traffic problems.The establishment of an intelligent urban traffic system is one of the important strategies to solve these problems.Traffic demand prediction is an indispensable part of intelligent transportation system.Accurate traffic demand prediction is of great significance in traffic management and urban planning.Due to the complex spatio-temporal characteristics of traffic demand,it is a challenging task.Existing methods for traffic demand prediction mainly focus on modeling the spatial correlation between geographically adjacent regions,and modeling the spatial correlation between distant regions through the similarity of some regional fixed features.However,due to these invariant regional features,the spatial correlations obtained from them are static.These methods all ignore that the spatiotemporal correlation will change with time,showing dynamic characteristics,such as passenger flow and short-term demand similarity between regions.Dynamic spatiotemporal correlation can reflect the travel status of residents in real time and is an important factor in accurately predicting demand.This paper studies the demand prediction task based on the deep learning method.The main contents are as follows.(1)Traffic demand prediction based on spatio-temporal dynamic multi-graph.To solve the above problems,this paper proposes Spatio-Temporal Dynamic Multi-Graph Attention Network(STDMG)for regional-level demand prediction.First,the longterm and short-term feature similarity between regions is encoded into multiple static graphs and dynamic graphs at each time step.Then,multiple graphs are fused to realize the input of graph information,and by modeling these graphs to capture spatial dependencies.Finally,a temporal attention module composed of ConvLSTM layers and attention layers is designed to capture the effects of adjacent spatiotemporal dependencies by incorporating global context information.This paper conducts four prediction experiments on two real public datasets,among which the root mean square error of the STDMG model is reduced by up to 5.8%compared to the benchmark algorithm,which verifies the effectiveness of STDMG in the task of traffic demand prediction.(2)Short-term demand prediction for ride-hailing in Hefei.There are certain differences between the original data collected in practical applications and the online public data sets,and the model is improved based on these differences.The original data of Hefei online ride-hailing includes order data and vehicle trajectory positioning data.Through data cleaning,collection,statistics and other methods,the division of regions is realized,the demand of each region at each time step is obtained,and POI data and flow direction data are introduced.It has been improved and applied to the demand prediction of ride-hailing in Hefei.By comparing with multiple models and with the unimproved STDMG model,the root mean square error of the improved model has been reduced by up to 3.65%,indicating that the improved model is more suitable for the short-term demand prediction task of Hefei online ride-hailing.