Research On Hotspots Identification And Prediction Based On Machine Learning For Cellular Networks

With the popularity of mobile multimedia services,mobile applications are becoming more and more abundant,and the data traffic demanded by mobile users has increased tremendously.As network quality requirements are critical to various data services,the telecom operators are driven to accelerate network performance optimization,so as to fulfill service and capacity demand of users in hotspot cells experiencing a significant amount of data traffic.To improve efficiency and quality of dynamic network optimization,it is vital to identify and predict potential hotspots in the network accurately and efficiently,so that timely adjustment and allocation of network resources can be achieved,the occurrence of network congestion can be reduced,and the network can be kept running smoothly.Therefore,this thesis takes traffic hotspots as the research object.Support Vector Machine(SVM)algorithm is proposed to identify traffic hotspots intelligently.Considering the time correlation of data traffic,Long Short-Term Memory(LSTM)model is adopted to predict potential traffic hotspots.These can serve as references and guidance for practical network management and optimization.In this thesis,based on the measured key performance indicators(KPI)data of the cellular network cells,time tread of traffic load about hotspots,the proportion of hotspot numbers across the time periods,and the frequency of occurrence are analyzed at first.It is found that the instantaneous traffic of hotspots has bursty characteristics,and the nature of hotspots changing with time is revealed.Secondly,the SVM machine learning algorithm is proposed to achieve intelligent identification of traffic hotspots.The Principal Component Analysis(PCA)method is considered to extract the key feature vectors of the original dataset for reducing the feature dimensions;linear kernel,polynomial kernel,radial basis kernel functions are utilizing to construct different SVM classifier models;and grid search technology is adopted to find the optimal parameter combinations for SVM with different kernel functions.Comparing and analyzing the hotspots identification results of the three kernel function SVM models with the optimal parameter settings.The numerical results indicate that SVM with radial basis kernel can achieve the best hotspots recognition accuracy,and it reflects the feasibility and effectiveness of the SVM algorithm in hotspots identification.In the end,an LSTM deep network model is designed to predict future traffic hotspots.This thesis establishes an LSTM neural network architecture consisting of two hidden LSTM layers and one fully connected layer,selects the appropriate activation function and optimization training method,inputs the preprocessing training data into the LSTM network for training,and outputs the classification result of forecasting whether the cell will be a hotspot in the next hour.During the training process,the four parameters of the number of training iterations,the number of neurons in the LSTM hidden layer,the regularized dropout probability,and the number of LSTM layers are controlled and adjusted.The prediction performance of the LSTM model is repeatedly optimized,and the impact of different parameter settings on the prediction results is analyzed.Compared with the traditional Feed-Forward Neural Network(FFNN),the proposed LSTM model has achieved good prediction results in the test set,which proves the ability of the LSTM network to memorize previous information and its superiority in processing time series data.