Dual-Step Transfer Learning-Based Prediction Model for Next-Generation Intelligent Cellular Networks

Deep learning-enabled cellular traffic modelling and prediction is an indispensable part of next-generation intelligent mobile networks. It can help proactive resource management operations in big data-driven cellular networks. Existing approaches usually adopt a well-known prediction model for univariate time series forecasting and require a huge amount of training data. In this work, we propose a dual-step transfer learning-based prediction model for multivariate spatiotemporal cellular traffic prediction. This framework groups the base stations into different clusters based on the traffic correlation. A base model is trained on the aggregated dataset of one cluster by using a combination of Recurrent Neural Network and Bidirectional Long-Short Term Memory (RNN-BLSTM). By using parameter-based transfer learning, the base model is provided to the other clusters as a starting point where it undergoes the process of fine-tuning on an intra-cluster aggregated dataset. After that, the obtained intra-cluster prediction model is transferred to all the base stations inside the cluster, where simple training is required with few parametric adjustments. We conduct extensive experiments on a real cellular traffic dataset and results show that the proposed dual-step transfer learning-based prediction model can achieve good prediction accuracy while reducing the data, time and computation requirements.