Intelligent RAN Power Saving using Balanced Model Training in Cellular Networks

optimizing power consumption of 5G systems and next generation technology deployments is a critical problem. It is essential that the solution for optimizing power consumption takes into account the tradeoffs with maintaining service level agreement (SLA). Mobile network operator (MNO) may have different priorities for the objectives of saving power and for maintaining SLA, which depends on factors, such as customer contract, location, time of day, type of traffic, etc. In this paper, we design an intelligent solution using switching cells on-off action to save power, using machine learning (ML)/ deep learning (DL) methods to forecast future traffic load. We firstly identify the problem of training imbalance in traffic load prediction due to data imbalance in real cellular networks, and MNO preferences for the competing objectives of saving power and SLA maintenance. We then propose a novel solution that incorporates Balancing Loss Function, which addresses the training imbalance problem. Compared with the performance of previous approaches such as Mean Square Error (MSE) minimization traffic forecast based methods, we demonstrate using network field data that our method is able to achieve upto 3X improvement in service quality outage, with fairly similar power savings.