| The continuous increase in urban vehicles has resulted in a growing trend of vehicle flows,exacerbating the issue of uneven distribution within urban spaces.Limited resources,such as space and roads,further compound the problem,leading to various challenges in urban development,including traffic congestion,parking difficulties,and road safety concerns.Effectively predicting and understanding the distribution of vehicle flows in urban regions is vital to address these issues.This thesis emphasizes the significance of accurate vehicle flow prediction and distribution analysis in urban areas.By enabling real-time traffic management,optimizing urban planning,and offering intelligent travel services,these insights serve as a foundation for enhancing urban traffic efficiency and ensuring a smoother transportation experience.In recent years,the proliferation of technologies like vehicle-mounted intelligent terminals and satellite positioning systems has facilitated the widespread collection and storage of vehicle trajectory data.This wealth of data provides valuable support for analyzing human travel behavior in urban regions and predicting vehicle flows.The dynamic nature of vehicle flow in urban areas,which varies both temporally and spatially,poses a significant challenge.While existing vehicle flow prediction methods primarily focus on capturing spatio-temporal correlations of vehicle travel,they often overlook the influence of various factors,including regional functions and weather,on the evolution of vehicle flow.Consequently,integrating the diverse spatio-temporal aspects of these factors and characterizing the functional semantic complexity of urban regions becomes crucial and challenging for accurate vehicle flow prediction.Existing methods struggle to capture the multiple spatial correlations of vehicle flow distribution within urban regions and the global temporal correlations of vehicle flow evolution across multiple dimensions.Furthermore,the data sparsity resulting from the uneven distribution of vehicles in urban regions further hampers the accurate prediction of vehicle flow.Addressing these challenges is essential to enhance the precision of vehicle flow prediction in urban regions.Based on a multitude of multi-source spatio-temporal data,including vehicle trajectories,urban regions,areas of interest,points of interest,weather conditions,holidays,and events,this thesis focuses on the diversity of spatio-temporal data,the complex functional semantics of urban regions,and the multiple spatio-temporal correlations in predicting vehicle flow evolution within urban regions.To address the challenges posed by key characteristics and sparsity of vehicle flow transfers,a series of innovative prediction models based on graph networks is proposed.These models are used to predict vehicle inflows,outflows,stay volume in urban regions,and vehicle flows between urban origin-destination(OD)regions,respectively.The main research work and specific contributions of this paper are as follows:(1)This thesis proposes a multi-source data fusion-based graph convolutional gated recurrent network(MDF-GCRN)to predict vehicle inflows and outflows in urban regions.MDF-GCRN integrates multi-source spatio-temporal data such as vehicle trajectories,urban regions,weather,holidays,and events,and learns spatio-temporal data diversity through graph convolutional gated recurrent networks and multi-head self-attention networks,respectively.The model establishes links between vehicle trajectory data and urban region data based on spatial proximity,enabling characterization of vehicle trip distribution.Graph convolutional gated recurrent networks capture temporal and spatial correlations of vehicle flow evolution,while a multi-head self-attention mechanism captures the diversity of external factors affecting vehicle flow.Experimental results on real multi-source spatio-temporal datasets demonstrate that MDF-GCRN significantly improves prediction accuracy,achieving a root mean square error improvement of 18.7%-20.9% and a mean absolute percentage error increase of 27.1%-31.3% compared to the current optimal baseline.(2)Aiming at the difficulty of parsing the functional semantic complexity of urban regions,this thesis proposes a semantic perception dual-view convolutional network(SPDCN)to predict the inflow and outflow of vehicles in urban regions.SPDCN first identifies urban functional regions and predicts vehicle flows within each region.It employs a multi-scale regional topic clustering algorithm to capture the functional semantic complexity of urban regions and identify urban functional regions,enabling characterization of vehicle travel distribution.The dual-view convolutional network models the spatio-temporal correlation of vehicle flow evolution from both the pixel perspective and the graph perspective.The pixel perspective models local spatio-temporal correlations within urban functional regions,while the graph perspective models traffic flow evolution at the regional level.The model combines global spatio-temporal correlations from both perspectives to predict future vehicle inflows and outflows in each urban functional region.Experimental results on vehicle trajectory and point-of-interest datasets demonstrate the superiority of SPDCN,achieving an average absolute error improvement of 43.6%-49.3% and a root mean square error improvement of 46.5%-61.1% compared to the current optimal baseline.(3)Aiming at the problem that it is difficult to capture the multiple spatio-temporal correlations of vehicle flow evolution,this thesis proposes a collaborative recurrent based on multi-graph dense convolution network(Co Re MDCN)to predict the vehicle stay volume in urban regions.Co Re MDCN models the evolution of vehicle stays and multiple spatiotemporal correlations between urban regions using a multi-view spatio-temporal graph.Densely connected blocks are introduced into a multi-graph convolutional network to comprehensively capture multiple spatial correlations.A collaborative recurrent network and an attention network are designed to capture serial time correlations of vehicle flow evolution and stay time correlations in urban regions.Experimental results on private vehicle trajectory datasets demonstrate the effectiveness of Co Re MDCN,achieving improvements of 14.6%-19.1% in coefficient of determination,17.2%-20.8% in explained variance score,11.8%-25.2% in mean absolute error,16.9%-18.3% in root mean square error,and 12.3%-18.2% in weighted average absolute percentage error,compared to the current optimal baseline.(4)Aiming at the sparsity of vehicle transfer flows,this thesis proposes a multi-graph generative adversarial network(MG-GAN)to predict vehicle OD flow between urban regions.MG-GAN solves the sparse traffic transfer records caused by the uneven distribution of vehicles in the urban OD region through the generative adversarial process.Specifically,a multi-graph convolutional gated recurrent network is used as a generative network to capture the multiple spatio-temporal correlations of vehicle OD flow evolution.Graph attention-based multi-graph convolutional networks as discriminative networks to personalize the attractiveness of urban regions.The generation network and the discriminative network improve the overall prediction accuracy through continuous iterative confrontation learning.The experimental results on the vehicle trajectory dataset show that.Compared with the current optimal baselines,the prediction accuracy of MG-GAN has increased by29.7%-47.6% in kullback-leibler divergence,and by 28.9%-47.9% in mean square error. |
