Master thesis presentation: Traffic Pattern Modelling for 5G Base Stations

As the number of network-connected devices continues to grow, the demand for accurate traffic models, which are essential for designing and optimizing 5G networks, increases as well. Currently, developers use statistical models, but with machine learning models rising in popularity, an idea has formed to explore the possibility of training a model on telemetry data. In this thesis, we explored uplink and downlink traffic from a 5G base station at native slot-level resolution (0.5 ms), using telemetry collected in an Ericsson test environment. We reconstructed consistent slot-level sequences from raw scheduling logs, and derived three event-centric representations: burst intensity (BI), inter-burst gaps (IBG), and a joint BI–IBG pair representation. We then trained and evaluated two neural generative approaches: a transformer based on the Chronos T5-style architecture, and an LSTM sequence model, and generated long synthetic sequences via autoregressive rollout. We finished with a detailed analysis, which shows that the original traffic is highly bursty, strongly imbalanced, and heavy-tailed, with clear regime shifts over time. While the models succeeded in learning marginal token distributions and some window level statistics, both architectures fail to capture the temporal dynamics needed for realistic slot-level traffic. Furthermore, we came to the conclusion that the main bottleneck lies not in the choice of model architecture but in the combination of the given constraints. 

The thesis work is done in collaboration with Ericsson AB.

Examiner: Maria Kihl