The Data Mining Based Load Analysis Of Base Station In TD-LTE Network

With the rapid development of mobile communications,the growth of data size increases explosively,which leads to the era of big data.On the one hand,a large number of data contains great.value which is as valuable as a gold mine.On the other hand,there are large challenges for the operators because of the surge in the number of users,the rise of base station load and the user experience need to be improved.In recent years,data mining and big data technology continue to explore huge value based on the data from all walks of life,attracting attentions of academics and Internet industries.In this thesis,a research is conducted on load analysis based on the latest data of Time Division Long Term Evolution(TD-LTE)network.The purpose is studying the rules and characteristics of traffic load,putting forward the corresponding strategy of cost,making load-balancing scheme and resource allocation proposal.The specific research content includes the following four aspects.In the study of the temporal and spatial distribution rules of the base station load,the data is preprocessed and the original load data is aggregated according to different time scales.For time domain,the periodicity,time autocorrelation,uplink and downlink,load difference are studied.As for the spatial domain,the differences of the entire network in different time are studied by using the spatial distribution of the Voronoi diagram of base stations.The original Moran index is optimized according to the Gauss kernel function and a further study on spatial autocorrelation of base station is made.Finally,a quantization index of load balancing is proposed,and the distribution of the index in the actual data is calculated.In the study of time series feature extraction of base station load and region type identification,the cross-correlation of the features of traffic load is studied,and the principal component analysis is adopted to reduce the dimension of the most relevant features.The extracted features are used for cluster analysis which leads to five different behavior patterns.Then the data set of these five types is divided into weekdays and weekends,and the load of different types of base stations on weekdays and weekends is studied.A reasonable dedicated billing strategy is proposed for operators to predict the type of base station.Formulating the strategy according to the time distribution of different types of base station.In the study of feature extraction and classification of high-risk base station,the index of load balancing is used to mark base stations with high risk and normal label.The samples are resampled and multivariate Gauss distribution stochastic noise are added to prevent risk of over-fitting.Then the feature engineering is conducted.The fusion idea in engineering is referenced to propose an optimized classification model.Gradient Boosting Decision Tree is used to combine and encode the features of base station,and the sparse discrete feature matrix after encoding is used as the input of Logistic Regression.The new model can improve the performance and maintain good generalization ability.In the study of the forecasting algorithm of base station load,the goal is trying to predict the load data of the base station itself only by using the time series data of the base station.The forecasting with the traditional time series analysis is conducted by dividing the series into three components trend,seasonal decomposition,trend decomposition and the residual.The wavelet decomposition is also used to decompose the series into high frequency and low frequency components.The root mean square error reveals the scheme of wavelet decomposition and time series analysis.The prediction error is reduced in the premise of performance which is a reference for operators to make energy saving strategy.