Traffic prediction with network slicing for 5G and beyond networks

This thesis proposes a Deep Neural Network-based method for predicting traffic volume that can guarantee that networks violate service-level agreements (SLAs) in small percentages against traffic demand. The proposed approach is multivariate, multi-step, and SLA-driven, as it incorporates 20 different radio access network (RAN) features, a custom feature set based on peak traffic hours, and mobility-based clustering to leverage spatiotemporal effects. In addition, different custom loss functions are proposed to ensure the determined SLA violation percentage while minimizing overprovisioning. Multi-step prediction up to 24 hours ahead is also performed, and the performance is compared with single-step prediction. The proposed method has been tested with two types of real and live mobile network traffic datasets, total and slice-based, on cells operating within a metropolitan area. The training architectures considered for the total traffic dataset are single-cell and multi-cell. However, the method focuses on single-cell scenarios with single-slice and multi-slice training architectures for slice-based traffic dataset. This thesis makes several contributions, including the analysis of RAN features, the custom feature set design, a custom loss function, and a parametric method to satisfy SLA constraints.