Research And Application Of Human Mobile Behavior Prediction Based On Deep Time Series Model

The study of the mobile trajectory of urban population,especially the daily flow of urban population,has become a hot part in the field of deep learning in recent years.Mobile trajectory can characterize human life rules such as commuting and work,as well as social preferences such as leisure and tourism,and it is also a visual indicator to monitor population mobility.From intelligent transportation,urban grid traffic prediction,urban planning and social stability maintenance,to mobile communication resource management,personalized recommendation systems and mobile medical services,the research on human mobility behavior prediction is of great significance.In this thesis,we mainly apply deep time series models to mine urban population travel patterns and periodic migration characteristics,especially optimizing two key problems in predicting human mobility behavior based on deep temporal models,namely,trajectory data is too long,sparse or redundant,and travel has high-order uncertainty and is different from person to person.In this thesis,we first propose a general method to classify user mobility preference based on mobility entropy and gyration radius.Then,the trajectory is clipped at different time intervals and the trajectory similarity is calculated according to the Frechet distance,so as to calculate the period more suitable for the user’s movement.Finally,the visit ranking algorithm is introduced,and the travel enthusiasm is quantified by Zipf s law,which corresponds to the stay time of the visit location attribute and the Point of Interest(POI),and the prediction module based on Attention Mechanism is designed.After making innovations in data processing and extracting travel patterns,this thesis also improves the structure of the deep time series model,and builds the key hyper parameters of the number of flashback states and spatio-temporal attenuation factor into the model for adaptive learning.And a multi-task prediction module is designed to add auxiliary prediction of the time and type of the next activity.Finally,this thesis applies front-end development and big data related technologies,and applies the existing data processing methods and trajectory prediction algorithms to the population mobility monitoring platform of Xuchang City,Henan Province,to help the construction of smart cities.