Hybrid Prediction Model for Mobile Data Traffic: A Cluster-level Approach

Mobile data consumption is rapidly growing following the ever-increasing bandwidth-hungry applications and improvements in network data rates. With the anticipated 5G right at the corner, operators are focusing on load-aware network dimensioning, optimization, and management, where traffic volume prediction plays a critical role. To this end, several researchers investigated different statistical and machine-learning models to exploit and predict the linear and nonlinear patterns that often arise due to the complexity of mobile networks and varying users' behaviors at different times and locations. In this paper, we propose a hybrid model composed of Double Seasonal ARIMA (D-SARIMA), which focuses on modeling the multi-seasonal nature of the data traffic and exploiting the residuals of DSARIMA via Long-Short Term Memory (LSTM)-based Networks. The residues contain the nonlinear component of the data. To incorporate the spatial dependency inherent in mobile data traffic collected from base stations, we used K-means clustering and considered the correlation among them. Experiments conducted with real-world data sets collected from 739 base stations for over four months, shows that our proposed hybrid model outperforms both D-SARIMA and LSTM models. The improvement emanates from capturing the double seasonality, non-linearity, and spatial dependency inherent in data traffic.