Research On Forecasting Algorithm And Application For Locator Data

With the rapid development of wireless communication system and popularity of smart phones,it has been becoming easier to obtain the user's location information.Massive LBS(Location Based Service)data has continuity in time and correlation in space,and can be used to analyze and predict.Location data can reflect the change of population density and thus to detect the crowd abnormal events.In this thesis,Tencent public location data is analyzed and preprocessed.The basic time series forecasting models are investigated.The multi-step prediction strategy and hierarchical prediction algorithms are studied and then the prediction model based on GBDT(Gradient Boosting Decision Tree)and LSTM(Long Short-Term Memory)are optimized separatelyThe main work and contributions of this thesis are:(1)GBDT model Optimization.The Moving Normalization mechanism based on GBDT is proposed and the features for GBDT are extracted.When the data distribution changes,it is difficult for the tree-based GBDT model to learn the data changes quickly.Therefore,this thesis puts forward the Moving Normalization mechanism which can normalize the input data and redistribute the prediction results to solve this problem.The experimental results show that the Moving Normalization mechanism in GBDT changes MAPE-NO(Mean Absolute Percentage Error of None-0 data)from 14.9%to 16.8%when the data distribution does not change,which is not significant,from 24.9%to 17.5%when the data distribution changed,which is significant.In addition,since it is difficult to automatically extract spatial features and abstract pattern features for GBDT model,time series related features suitable for GBDT are designed.Experiments show that using the time series features and spatial features and clustering information can change MAPE-NO from 18.5%to 14.9%relative to using original data only.(2)LSTM model Optimization,bidirectional LSTM structure based on LSTM model and Batch Normalization mechanism are investigated.The experimental results show that the Bi-LSTM(Bi-directional LSTM)changes MAPE-NO from 14.9%to 18.4%.(3)Implementation of the application system based on LBS data.The application system is designed,which includes data preprocessing,feature extraction,LBS data prediction,abnormal event monitoring and security warning.The hierarchical prediction strategy of LBS data is studied,the LBS data prediction problem is modeled as hierarchical data prediction problem,and the hierarchical consistent error loss function is designed.The multi-step prediction strategy is investigated to transform the single-step prediction results of the model into multi-step prediction for practical application.