Prediction Of Staying-crowd Flow And OD Flow Based On Mobile Signaling Data

With the development of Chinese modern cities and the popularity of smart phones,more and more cellular base stations are deployed in the city,and the scale of mobile signaling data is increasing rapidly.By analysing the signaling data,we can monitor and guide the urban crowd flow.Different from the traditional method which divides the city area into grids,In this paper,we predict the staying-crowd flow and the OD flow of area covered by base station.So the flow can be effectively specified and located.It not only helps the base station to save energy and schedule the resources,but also greatly improves the utilization of urban resources,urban traffic efficiency and the city's ability to respond to sudden disasters.This paper mainly uses graph convolutional Network and graph embedding method to predict crowd flow.We use Changchun mobile signaling data to simulate urban crowd flow,and make statistical analysis on base station graph structure and the crowd flow in the city.The main contents of this paper are as follows:First of all is the data cleaning and processing.Then we use Changchun mobile signaling data,Changchun POI data,base station location data to restore the trace of residents and calculate staying-crowd flow and OD flow of base stations.Secondly,since the signal base station does not have the same natural road graph structure as the urban road,it is a challenging work to perform graph convolution and graph embedding on the base stations.In this paper,a graph fusion method is used to model the traffic transfer relationship,geographical location distance,and urban functional meaning of the base station.Specifically,we use the number of flow records between base stations to build the adjacency matrix to indicate flow situation,use the distance between base stations to build the distance adjacency matrix,and then matches the urban POI data to the base stations to calculate the jaccard correlation coefficient of the POI,to generate the jaccard adjacency matrix which indicate the similarity of regional function.By learning the weights of different graph structures through a set of parameters,finally we get the graph structures between base stations.Then,we connect the time graph convolution model after the graph fusion module,which can capture the time-space dependence of base station traffic by using the GCN and the GRU.Taking the time series features extracted from the base station crowd flow as the features of the graph nodes,the features are extracted through the GCN module,and then input into the gate recurrent unit to make short-term prediction of the base station staying-crowd flow.The experiment shows that the performance of the model constructed in this paper is better than the traditional time series model,and the graph fusion module can effectively learn the relationship between different graph structures of base stations and improve the performance of staying-crowd flow prediction.Finally,we use the graph structure generated by the graph fusion module to make graph embedding.By using node2 vec,Each base station is represented by a embedding vectors,which indicate the implied mode of traffic transfer between base stations.In addition,the detailed feature engineering of OD traffic transfer between base stations is constructed by using graph theory,statistics and time series related knowledge.In order to verify the effectiveness of the graph embedding method,we use Light GBM and random forest to build a OD flow prediction model,which uses the graph embedding vectors,POI distribution features,historical flow features,graph structure features.The experiment proved that the graph structure generated by graph fusion module is better than other methods in graph embedding.