TL-ConvLSTM: A Transfer-Learning-Based Convolutional LSTM to Identify and Forecast Traffic in the NextG Environments

Abstract

Forecasting and categorizing cellular traffic flows and their types are essential functions in intelligent network systems to ensure efficient network optimization. The ever-evolving nature of 5G networks results in fluctuations in traffic patterns over time, leading to a phenomenon known as model drift. Consequently, accurately predicting and identifying cellular traffic patterns becomes a complex task. To tackle this challenge, this article introduces an innovative approach called TL-ConvLSTM, which combines transfer learning with convolutional long short-term memory (ConvLSTM) to effectively combat model drift and provide precise forecasting and recognition of cellular traffic within the network. To accomplish this, we initiate the training of TL-ConvLSTM by estimating its parameters from the source domain. We then employ the Kolmogorov–Smirnov method to adapt the model within the target domain, fine tuning its weights. To improve the precision of this model adaptation, we systematically explore optimal learning windows. This exploration includes adjusting window size for time-series data and feature dimensions to capture dynamic traffic patterns in a 5G environment. Furthermore, we make use of the Amarisoft 5G testbed in our lab to create a 12-day time-series dataset. This dataset includes various features related to traffic flows and their patterns. We showcase the effectiveness of our approach through a set of experiments.