Machine learning mechanisms for forecasting the performance of a mobile network

• Autor(a) principal: Pousa, Ricardo José Baptista
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/38984
• topic: Network monitoring; Time-series; Machine learning; Network state prediction; Clustering; Forecasting

With the growth in bandwidth and service complexity in Mobile Networks, in conjunction with its increasing demand, it is expected to occur a paradigm shift in the way these operate: from the increase in control provided by SDNs (Software De ned Networks) to the wide range of proposed services. Network operators will have to accompany the expected growth with more intelligent and automated methods of network monitoring and management, that in turn will allow for its further optimization. These networks are expected to generate a large amount of data as Time- Series, and therefore it is required to build systems that can e ciently work with them. Depending on the location of the eNBs, the studied KPIs may exhibit rather varying dynamic behaviour, therefore, data was divided into clusters with similar behaviour using K-means clustering. Clustering was applied with different features extracted from the raw time-series data of the KPI ERAB Attempts. Techniques such as Discrete and Continuous Wavelet Transforms, Discrete Fourier Transform and Time Series Decomposition (seasonal, trend and remainder components) were used as feature engineering methods. In this dissertation two Machine Learning (ML) methods are applied to predict the mobile networks' eNBs (eNodeB) states, namely Random Forest and Ridge Regression algorithms. The goal is to identify the Critical and the Degradation states in two protocols - ERAB and RRC. Further to that, Root Cause Analysis was done to estimate the major features responsible for the identi ed KPIs states. The nal task of this thesis was to compare 3 time-series forecasting ML approaches to predict the next hour value of the studied KPI (ERAB Attempts). Extreme Gradient Boosting (XGBoost), Feedforward Neural Network (FFNN) and pretrained FFNN (transfer learning) were applied to the clustered data (sub-datasets), whereas the rst two were applied only to non-clustered data. In forecasting it was observed that the clustering led to improvements in the forecasting accuracy, and simultaneously decreased the quantity of models needed to cover every node of a Mobile Network.