Resource demand prediction for network slices in 5G using ML enhanced with network models

The new technologies introduced by 5G, such as network slicing, will improve the capabilities of Vehicle-to-Vehicle (V2V) communications, enabling the introduction of a new range of services and new forms of Vehicle-to-Everything (V2X) interactions. In order to deploy these V2X services and the network slices they are associated with over the 5G network while ensuring Quality of Service (QoS), intelligent and proactive network resource managers and orchestrators (RMOs) need to be developed. The ability to forecast the slice resource demand can significantly increase the proactivity of these RMOs. ML-based resource demand predictors (RDPs) are commonly integrated with RMOs to provide accurate forecasts of the slice resource demands in V2X use cases. However, prediction errors are still common, causing the RMOs to reallocate resources to the slices sub-optimally. When an RDP underestimates the resource demand, i.e. predicts less demand than expected, the impact is much more severe for the infrastructure providers (InPs) and service providers (SPs) than when it overestimates the demand. Also, the impact of this misprediction is also different for each InP/SP, for which it is necessary for RDPs to also consider this difference. In view of this, we introduce a new approach that makes ML-based RDPs aware of the asymmetry of misprediction and their dependence to a specific network model, making their forecasts more useful for RMOs. This approach enhances the design of RDPs by embedding within them knowledge of the underlying 5G network and of the relationship between resource demand, resource allocation and service/network performance. We refer to our approach as Network-Aware Loss for Demand Prediction (NALDEP), and it improves the prediction quality by 73.3% and 41.0% with respect to accuracy-based and other state-of-the-art predictors, respectively.