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

Traffic modeling and prediction are indispensable to future extensive data-driven automated intelligent cellular networks. It contributes to proactive and autonomic network control operations within cellular networks. Current methodologies typically rely on established prediction models designed for univariate and multivariate time series forecasting. However, these approaches often demand a substantial volume of training data and extensive computational resources for prediction model training. In this study, we introduce a dual-step transfer learning (DSTL)-based prediction model specifically designed for the prediction of multivariate spatio-temporal cellular traffic. This technique involves the categorization of gNodeBs (gNBs) into distinct clusters based on their traffic pattern correlations. Instead of training the prediction model individually on each gNB, a base model is trained on the aggregated dataset of all the gNBs within a base cluster using a combination of recurrent neural network (RNN) and bidirectional long-short term memory (RNN-BLSTM) network. In the first-step transfer learning (TL), the base model is provided to the gNBs within the base cluster and to the other clusters, where it undergoes the process of fine-tuning the intra-cluster aggregated dataset. Once the model is trained on the aggregated dataset within each cluster, it is provided to the gNBs within the respective cluster in the second-step TL. The model received by each gNB through the proposed DSTL technique either necessitates minimal fine-tuning or, in some cases, requires no further adjustment. We conduct extensive experiments on a real-world Telecom Italia cellular traffic dataset. The results demonstrate that the proposed DSTL-based prediction model achieves a mean absolute percentage error of 2.97%, 9.85%, and 9.73% in predicting spatio-temporal Internet, calling, and messaging traffic, respectively, while utilizing less computational resources and requiring less training time than traditional model training and TL techniques.