A cluster-based approach for cellular traffic prediction with machine learning methods

• Main Author: Correia, Daniel Vala
• Publication Date: 2023
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10773/41694
• topic: 5G; Cellular traffic prediction; Clustering; KPI prediction; Machine learning; Recurrent neural networks; Smart urban mobility; Time series

Over the last decade, technologies like Fourth Generation of Cellular Communications (4G) and Cloud computing have profoundly transformed our way of living and communicating, increasing cellular traffic by tenfold every five years. Technologies like Fifth Generation of Cellular Communications (5G) and Internet of Things (IoT) are expected to continue this increasing trend. To handle such large and complex networks, the community strives to achieve self-management, intelligent Autonomous Networks (AN). With accurate traffic prediction, we can make autonomous decisions in planning, management, resource allocation, etc. This dissertation devotes itself to implementing and testing a cellular traffic predictive system of the state of the communications network, more specifically a predictive system of Key Performance Indicators (KPIs), from a major Internet Service Provider (ISP) in Portugal. The selected KPIs are highly correlated with the cellular network traffic. This system uses a grid-based clustering algorithm to group cells based on their spatial dependencies. Then, several methods are used to forecast future values and the results are compared. The persistence model is used as a baseline, the Seasonal Autoregressive Integrated Moving Average (SARIMA) is used as the classic statistical method, and the Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) are used as machine learning methods. The results showed that GRU outperformed the SARIMA method and had a slight improvement compared to the LSTM method. Lastly, a study is done on the developed model (based on a cluster) when forecasting different clusters without being retrained, and when trained with data from the newly selected cluster. The results achieved are promising when forecasting other similar behavior clusters.