Study On The Citywide Crowd Flow Prediction Based On Spatial Temporal Big Data

Modeling and prediction of citywide crowd flow can provide scientific insights into the distribution of urban population,patterns in resident activities,optimization of public resource allocation,location choices for commercial facilities,and contingency plans for public safety.In the era of big data,the introduction of spatiotemporal big data and deep learning technologies brought new opportunities and challenges to the prediction of citywide crowd flow.Existing research methods lack the process of modeling the dynamic spatiotemporal characteristics of the data towards the spatial structure of irregular urban regions and do not have the ability to predict both the flow and direction of crowd flow.To change the status quo in research,this paper explores the combination of spatiotemporal big data and deep learning algorithms to model crowd flow among arbitrary urban regions at a fine scale.Here are the designs:(1)This paper presents a method for modeling crowd flow in urban irregular regions.Firstly,based on urban bus route data,this paper innovatively merges morphological geometry algorithm and means clustering algorithm to divide the city into a number of regions of any shape.Then,a flow-field generation algorithm is proposed to aggregate large amounts of low-value density trajectory data into low-volume,high-value density spatiotemporal flow-field data.Finally,the spatiotemporal characteristics of citywide crowd flow are analyzed from the crowd flow-field data,and formulaic definitions of the rotation,flux,divergence and ring flow of the regional crowd flow field are given.(2)In this paper,an adaptive spatiotemporal graph convolutional network is proposed for predicting crowd flow between irregular urban regions.A specially designed block is used to dynamically generate Laplace matrices adapted to capture spatial correlations of input spatiotemporal graph sequences,solving the problem that Laplace matrices predefined based on interregional distance cannot dynamically track the time-varying spatial connectivity state of regions.The input spatiotemporal graph sequence is constructed using a variety of time-dimensional sampling methods,and the modularly designed spatiotemporal graph convolution component is used to model the spatiotemporal features of the data to achieve accurate prediction of the crowd flow between irregular regions in the city.Experiments on the DDCD have shown that the method proposed in this paper can accurately divide the localized area of the Second Ring of Chengdu into 64 irregular regions.The proposed flow field generation algorithm generates a total of 275 MB of citywide crowd flow-field data,which is only 0.13% of the original data volume,greatly reducing data storage consumption.The experimental comparison found that the performance of the model presented in this paper was better than that of the four baseline models,with the mean absolute error reduction of 28.7% and the root mean square error reduction of 37.9%.Network variants with different inputs and different spatiotemporal graph convolution components perform differently,with the best combination achieving nearly 40% model performance improvement.This paper designs a prototype cloud service system for urban crowd flow prediction to verify the feasibility of this research method from the application level.This paper has 57 pictures,8 tables,and 66 references.