Study On The Method Of Active Population Prediction With Micro Spatiotemporal Granularity Based On Mobile Phone Signaling Data And Integrated LSTM And CA

The micro spatiotemporal granularity represents a relative concept in this thesis,which means that the temporal and spatial granularity are more refined than traditional research.Active population prediction can be understood as a type of population prediction.It refers specifically to predicting the number of people that appear in a given time and space,such as forecasting the number of tourists in a theme park at an hour in the future.The prediction of active population at the micro spatiotemporal granularity has practical significance for refined urban management and response to emergency disasters.In the past,due to the limitation of collecting information,methods of prediction were mostly based on censuses or statistical yearbooks.The measurement was relatively macro,and it was difficult to fully characterize the spatiotemporal changes of the active population.With the advent of the information age,the emergence of new methods for collecting data with fine-grained and complete samples,such as mobile phone signaling data,provides rich materials and new solutions for research.However,traditional methods of population prediction are not good at processing such large-scale and fine-grained data,and existing scholars are less able to deeply explore the internal characteristics from both the time and space perspectives.Combining the spatiotemporal characteristics of the data itself,and on the basis of drawing on existing research,in order to achieve a more precise prediction of the active population at the micro spatiotemporal granularity,this thesis proposes an integrated Long Short-Term Memory(LSTM)and Cellular Automata(CA)model,and uses the mobile phone signaling data collected by China Unicom in Chongming District of Shanghai as an example to verify the effectiveness of the method and its applicability in different scenarios.This method spatially divides the study area into regular grid cells.Then it utilizes the interaction mechanism of neighboring cells provided by the CA to reflect the movement characteristics of the active population,and uses the LSTM to realize active population prediction based on grid cells in the time series.The research in this thesis is mainly divided into three parts:(1)introduce and summarize the existing methods of population prediction and research on mobile phone signaling data;(2)aim at the research goals and objects,then propose a data-driven prediction method of active population which integrating LSTM and CA;(3)use the mobile phone signaling data collected by China Unicom in Chongming District of Shanghai to build a prediction model of China Unicom's users in this region,and set a variety of scenarios to evaluate the prediction results quantitatively.Combined with the case data,the results show that the proposed integrated model can achieve a good result.Among them,the model trained by three consecutive days(ordinary dates),granularity of one hour,and the road network is used to calculate weights of neighborhood cells,is the best,and its Mean Absolute Error(MAE)is less than 1,which means that the model has good applicability in the prediction of micro spatiotemporal granularity of active population.The innovation of this thesis is different from the previous related research in that by integrating two models of LSTM and CA,it combines the two dimensions of time and space to fully exploit the short-term and spatial distribution characteristics of the active population,and draws the prediction results considering the spatial differences.In particular,this thesis also proposes an algorithm for merging information of the road network to determine the cell weights in the CA,which can further improve the accuracy of prediction.