Research On The Computation And Prediction Of Citywide Crowd Flows

Forecasting the crowd flows in urban areas accurately will provide effective decision-making support for governments and management departments to allocate urban resources rationally,improve the citizens' travel experience and eliminate urban safety dangers.The prediction of citywide crowd flows is a challenging problem.First,regional crowd flows data are spatio-temporal data,with strong dependencies on both spatial and temporal dimensions.At the same time,crowd flows are also affected by many external factors.Most traditional methods of the crowd flows pattern analysis often take a single region as the analysis object.It is difficult to make a comprehensive and accurate estimation of the global state of crowd flows only by considering the local spatio-temporal information of the regional crowd flows.Therefore,this paper gives a full consideration on the global factors which affect the crowd flows of multiple regions,to model and forecast the citywide crowd flows.This paper divides a city into multiple regions according to the gridding method.For any time interval and any region,we use the user's trajectory data to represent and compute its crowd flow,thus the entire city crowd flows can be modeled as tensor data,which contain both temporal and spatial dependencies.For research on crowd flow forecasting in grid urban region,this paper proposes a deep-learning based approach,Spatio-Temporal Recurrent Convolutional Networks(STRCNs),to uniformly model various relevant factors which affect the crowd flows properly and comprehensively,thus it can predict the population inflows and outflows in all regions of a city at multiple time intervals.Aiming at the problem of spatio-temporal dependencies in the prediction of crowd flows in urban region,STRCNs use Closeness captures the instantaneous characteristics of crowd flow changes;Daily Influence detects the daily regularity;Weekly Influence reacts the weekly patterns;we combines CNN and LSTM to learn both spatial and temporal dependencies in each component,which CNN describes the spatial dependencies,and LSTM characterizes the temporal dependencies.Aiming at the problem of external factors description in the prediction of crowd flows in urban region.The external influence factors are extracted by STRCNs using External Influence component,this component is a two-layer fully connected neural network.For the output of the four STRCNs components,STRCNs assign different weights to different components,and merge all the outputs of the four parts together.Finally,we validate our proposed STRCNs model on two real datasets,MobileBJ and TaxiBJ respectively.Experimental results demonstrate that STRCNs outperforms classical time series methods and other existing deep-learning based prediction approaches.