Flow Prediction Based On Deep Neural Network

The prediction of flow plays an important role in smart cities,which mainly includes the prediction of regional flow and the prediction of POI flow.Regional flow describes the flow of groups in space,which is important in urban computing,traffic distribution and resource allocation.Accurate prediction of regional flow is a challenging problem.We need to capture both spatial and temporal patterns.At the same time,the existing scheme is to divide regions into the grid,and explore the flow relationship between grids.But the actual road network isn't a regular grid shape.It is difficult to apply to all scenarios by simply using grid division without considering the geographical characteristics of reality.Therefore,considering the geographical characteristics of reality,this paper uses graph divide the region.We creatively constructed the regional geographic adjacent relation graph and regional flow graph,and explored the spatial and temporal relation of regional flow by using the Graph Convolutional Recurrent Network.We chose the public dataset of NYC-taxi for experiments.Through the experimental analysis,the method is better than other methods and improves the prediction accuracy of regional flow.The RMSE was reduced by 0.05.The prediction of POI flow can help location-based services,help visitors plan their trips,and help shop manager to guide their business plans.How to predict the future POI flow is a worthy problem.Compared to the modeling of the regional flow,the geographical adjacent relation between POIs is difficult to define and represent,and the POI have the characteristics of functional types,so it is difficult to directly model the flow of groups between the POI.The researchers used a time-series method and a convolutional neural network to capture its temporal patterns.However,the crowd flow of POI does not always show temporal pattern,and may appear abnormal fluctuations.Therefore,this paper comprehensively considers the characteristics of historical flow and the characteristics of abnormal fluctuations,and uses the long short term memory network to model the characteristics of historical traffic.In addition,considering the influence of regional flow variation and the attraction of POI,we model the abnormal fluctuation.Finally predicting POI flow by integration.We collected a large number of check-in data to construct the dataset of POI,and carried out experiments.The experimental results show that the method presented in this paper has a significant improvement compared with the existing method.The RMSE was reduced by 24.5%compared with the original best algorithm.