RAN Congestion Prediction for Slice Automation

Network slicing will enable 5G network operators to offer extremely diverse set of services over a shared physical network infrastructure. Automation of network slice lifecycle management will play a key role in realizing the full potential of 5G and network slicing. Such automation requires reliable and fine grained prediction of network Key Performance Indicators (KPIs) with short time scope. This will enable automatic network slice reconfiguration and optimization actions to be taken pre-emptively, to avoid anticipated degradation network KPIs and Service Quality Indicators (SQI). Within mobile Radio Access Networks (RAN) one of fundamental KPIs is RAN congestion level, which together with other KPIs like wireless channel quality enables computation of service critical KPIs like network latency and throughput. In this paper, we define a new metric to measure RAN congestion. We propose a new time series prediction model for RAN congestion, based on Long Short Term Memory neural networks. We use experimental LTE RAN system to collect and evaluate the prediction model. Finally, we show that our model achieves over 90% prediction accuracy.