AI-Assisted Network Traffic Prediction Without Warm-Up Periods

Network traffic prediction in cellular networks improves reliability and efficiency of network resource use via proactive network management schemes. To this end, future traffic arrivals are anticipated via machine learning (ML)-based network traffic predictions based on historical network traffic data. Current literature on ML-based network traffic predictions employs warm-up periods, which are the required duration traffic flows are observed to make meaningful predictions. However, most flows are shorter than the warm-up period. This paper proposes a residual neural network (ResNet) architecture for individual network flow predictions, based on a deep-learning approach that removes the required warm-up period seen in other proposed methods. The ResNet architecture demonstrates the ability to accurately predict the magnitude of packet count, size, and duration of flows using only the information available at the arrival of the first packet such as IP addresses and utilized transport-layer protocols. The results indicate that the proposed method is able to predict the order of magnitude of individual flow characteristics with over 80% accuracy, outperforming traditional ML methods such as linear regression and decision trees.