Research On Forecasting Method Of Urban Area Crowd Flow Based On Deep Learning

Forecasting the flow of crowd plays an important role in urban traffic planning management and urban public safety.Accuracy is a challenge for predicting the flow of crowd in a region.On the one hand,the citywide crowd flow data is high-dimensional,and the original high-dimensional data usually contains redundant information,which will cause errors to the prediction results and reduce the accuracy in practical application.Most of the existing prediction models do not consider the impact of high-dimensional crowd flow data on the prediction accuracy and algorithm efficiency,and the network model has a complex structure and a large number of parameters,so training network needs to consume a huge cost.On the other hand,the prediction of crowd flow is influenced by many complex factors,including spatial structure correlation,dynamic temporal dependencies,and external(such as weather,holidays,and events).In order to resolve the above problems,this paper proposes a new prediction method of urban region crowd flow based on deep learning.Aiming at the impact of high-dimensional crowd flow data and spatio-temporal factors on model prediction accuracy and algorithm efficiency,this paper proposes a deep bottleneck residual network crowd flow prediction model based on BRBM(BRBM-B-Res Net).The model mainly consists of two parts: data reconstruction mechanism and collaborative prediction mechanism.The data reconstruction mechanism mainly uses BRBM to reduce dimensions and reconstruct high-dimensional crowd flow data.The collaborative prediction mechanism is mainly composed of spatio-temporal prediction sub-model and external auxiliary prediction sub-model.The spatio-temporal prediction sub-model uses the separated bottleneck residual network to model the crowd flow data into three time dimensions(hours,days and weeks),and in this way,the spatial and temporal dependence of crowd flow between different regions can be mined.The external auxiliary prediction sub-model uses a two layers fully-connected network to process external factor data,thereby assisting in predicting urban crowd flow.In ST-Res Net,the network structure with shutcut connection is easy to loss some feature information flow and hard to extract fine-grained time feature information.To solve that problem.This paper proposes a prediction model of densely connection network based on dilated convolution(ST-DDN).The model consists of three parts: the spatial prediction part,the auxiliary prediction part and the time correlation part.The spatial prediction part uses dilated convolution to improve the densely connected network,and models the spatial correlations on different levels between urban regions.In the auxiliary prediction part,a fully connected network is used to process the external factor data.The time correlation part adopts the SE-LSTM module to learn more fine-grained time feature information in crowd flow data and external factors data.In this paper,two datasets Taxi BJ and Bike NYC were used to verify the proposed models.Experimental results show that the BRBM-B-Res Net model and ST-DDN model not only effectively improve the prediction accuracy,but also have better performance than other existing baseline models.