Research On Intelligent Prediction Methods For Urban Traffic Based On Multi-Source Spatio-Temporal Data

In recent years,the rapid growth in residents’ travel needs has consistently outpaced the development of urban transportation,resulting in increasingly complex traffic conditions and worsening traffic congestion.Traditional urban transportation management methods can no longer meet the needs of smart city fine-grained traffic management and personalized travel demand of residents.At the same time,the widespread application of Internet of Things(IoT)technology has brought a large amount of multi-source spatio-temporal data,providing important support for the fine-grained modeling and accurate prediction of urban population and vehicle movement patterns.Conducting research on intelligent prediction methods for urban traffic based on multi-source spatio-temporal data can not only help governments efficiently dispatch traffic capacity,reduce the risk of emergencies and alleviate traffic congestion,but also provide reasonable travel advice and improve travel experience for the public.In this paper,we conduct research on intelligent prediction methods for urban transportation based on multi-source spatio-temporal data,focusing on the city roads,bus lines,and individual points.We effectively address technical challenges such as the decrease in spatio-temporal continuity caused by unexpected events,the weakening spatio-temporal correlation caused by prediction granularity,and the absence spatio-temporal periodicity caused by sparse data sampling.The main innovations and contributions of the paper are as follows:(1)Condition-aware traffic prediction method for urban road.For conditionaware traffic prediction,various factors,such as abrupt change in traffic patterns,dynamic spatial dependencies and sparsity of abnormal events,poses challenges in the prediction process.To adapt both normal and abnormal conditions,we propose a Multi-task Multi-range Multi-subgraph Attention Network(M3AN),a novel deep learning model to explicitly model the impacts of abnormal events for condition-aware traffic prediction.Firstly,this method explicitly models node features by constructing different subgraphs to effectively capture the impact of sudden sparse abnormal events on urban traffic flow.Secondly,this method uses a multi-task fusion module to jointly learn the traffic flow of road segments and intersections,which can better describe the interaction between different types of nodes and enhance the model’s ability to capture complex spatio-temporal correlations.Thirdly,this method builds a multi-scale attention module with low computational complexity to automatically learn the impact of abnormal events on traffic flow and improve robustness.Experimental results on two real urban traffic datasets from Tai’an City and the San Francisco Bay Area demonstrate that the effectiveness of M3AN.Compared to the state-of-the-art method,it achieves an average absolute error reduction of 9.21%and 40.69%in the respective datasets.(2)Service-level passenger flow prediction method for public transportation.For the multi-step service-level passenger flow forecasting,various complicated and dynamic factors,such as inter-station,inter-line and inter-service spatial-temporal dependencies,are not effectively modeled by existing methods.To address these challenges,we propose a spatio-temporal hashing multigraph convolution network(ST-HMGCN),a novel deep learning method to achieve service-level passenger flow prediction.Firstly,this method constructs two types of subgraphs from perspectives of physical adjacency and semantic similarity to explicitly capture spatial-temporal dependencies among bus stations/lines,and integrates the inter-service temporal correlations to achieve the service-level bus passenger flow prediction.Secondly,this method utilizes the hashing graph convolution to extract the dynamic spatial correlations among graph nodes.Thirdly,this method uses a temporal-attention block with residual connections to model the non-linear temporal correlations between different time intervals of each station,which significantly reduces the error propagation among prediction time steps.Experimental results on a large-scale bus operation dataset from Jinan City demonstrate the effectiveness of ST-HMGCN.Compared to the state-of-the-art methods,it achieves an average absolute error reduction of 8.25%and 8.51%in line-level passenger flow and bus service flow prediction,respectively.Additionally,it shows an average absolute error reduction of 13.37%in service-level passenger flow prediction.(3)Proposed an OD-Prophet model for predicting personalized travel ODs.For the individual Origin-Destination prediction of next trip,various factors,such as uneven and sparse location sampling,time shifting of periodical mobility and diverse individual travel patterns,makes it more difficult to model the nonlinear transition regularities between trajectories and future trip OD.To tackle these challenges,OD-Prophet,a contrastive learning-based multi-view attention network is developed for individual OD prediction of next trip.Firstly,this method constructs a multi-view spatio-temporal embedding module to obtain personalized representations with enhanced group information with user ID,time,location,and individual/group OD travel patterns.Secondly,this method introduces a correlation-enhanced hybrid attention mechanism to alleviate the time shifting of periodical mobility.Thirdly,this method uses a supervised contrastive learning framework to generate high-quality representations from sparse trajectories and improve the generalization capability.Experimental results on large-scale datasets collected from Shenzhen and Xiamen cities demonstrate that the effectiveness of OD-Prophet.Compared to the state-ofthe-art methods,it achieves an accuracy improvement of 20.52%and 12.87%in origin prediction,and 110.81%and 98.07%in destination prediction,respectively.Finally,based on above methods,this paper implements a verification system of a smart urban traffic prediction based on multi-source spatio-temporal data,possessing a set of traffic prediction,traffic recommendation,and traffic situation display functions for three different scales of road traffic,public transportation,and personal travel,validating the advanced and feasible nature of the proposed methods.The outcome of this research can be widely applied to practical traffic scenarios,providing strong support for fine-grained management of urban traffic and personalized travel services.