Citywide Crowd Flow Prediction Based On Deep Networks With Spatiotemporal Attention

A city is generally the economic,political and cultural center of the region it located.In the city,The characteristics of people's daily going-out demands are diversified with high frequency.With the well-developed transportation,people can go out for commuting,shopping,entertainment conveniently.However,the crowded public transportation,dense crowds in public places and road traffic congestion remain the most complained problems.Nowadays,with real-time going-out tips of ‘crowdedness' and ‘traffic index',people can choose a more comfortable transportation mode or change the going-out plans in time.Besides,the measures for public safety and road traffic control are also more efficient.However,the indicators indicate those states in the current time rather than the future,which means the effect is relatively limited.Therefore,real-time prediction of citywide crowd flows has become a research hotspot.The existing studies focus on dividing the city into grids and predicting the next period of crowd flow in each grid area.Considering the rapid changes in citywide crowd flow,the measures for public safety or traffic control require response time and the locations of the peak of crowd flow.In this paper,we study the multi-step prediction problem of citywide crowd flow.Specifically,based on deep learning technology,this theis focuses on predicting the station-level citywide crowd flow(SLCCF),and study how to combine the geographical information of stations with deep neural networks to achieve accurate prediction.The main work and innovations are as follows:Firstly,a prediction model based on spatiotemporal attention mechanism is proposed for multi-step prediction of citywide crowd flow in a grid-divided city.Unlike most exist-ing studies using recurrent neural networks to iteratively predicting the multi periods,the model consists of multi layers based on the spatiotemporal attention mechanism,called ST-Attn.Without recurrent neural network units,ST-Attn adopts the multi-step directly output strategy.The spatiotemporal attention mechanism layers mix attention mecha-nisms from both time domains and space domains.The layers directly model the spa-tiotemporal correlation between each predicted grid area and time period and the other regions and time periods on the change of crowd flow.In addition,for multi-step pre-diction,the problem of the weakening spatiotemporal correlation between the subsequent predicted periods and the known observation period in terms of changes in crowd flow,ST-Attn introduces the spatiotemporal kernel density estimation.The rough prediction results of the prediction,thereby enhances the prediction performance for the later period to be predicted.Secondly,to deal with such issues,a next-step prediction model based on spatiotem-poral U-shaped network and a multi-step prediction model based on hierarchical spa-tiotemporal attention mechanism for SLCCF are proposed,called ST-Unet and ST-HAttn respectively.The existing convolutional neural networks are not suitable for irregular data domains.Thus,ST-Unet and ST-HAttn emphasize stations' spatial dependence by inte-grating the crowd flow information from neighboring stations and the cluster it belongs to after hierarchical clustering.Convolutional layers based the road networks where the stations lie are proposed.The notable contribution is that ST-HAttn performs attention mechanisms(AM)in two ways: 1)implementing AM at both station level and regional level? 2)implementing AM to explicitly model the pairwise correlations of station-region instead of station-station.The intuition is to alleviate the negative impact on Ms-SLCFP due to the fluctuation of the crowd flow at the station level.Thirdly,we look into the problem of ‘underestimation',which occurrs in the pre-diction of citywide crowd flow during those extremely peak periods,called burst-flow following.A model based on graph convolutional neural network is proposed to identify and re-estimate these periods with burst-flow.Specifically,the prediction model is based on the classification and regression framework.The burst-flow discriminant model is to predict whether there would be a burst flow period at the station.The burst flow regression model is to predict the exact value of crowd flow.The combination of both improves the accuracy of burst-flow prediction.Further,to deal with the problem of data imbalance,the method of adjusting the weight of category loss is adopted.Through the collaborative training of the classification network and the regression network,the prediction accuracy and discrimination accuracy of the burst-flow are improved on the whole.