Cell Tower and Cell Phone LTE Traffic Pattern Prediction with the Gated Recurrent Unit Neural Network

Abstract

We apply the gated recurrent unit (GRU) neural network model to the task of network traffic pattern prediction. We consider both cell tower and cell phone traffic patterns. Here, a cell tower's traffic pattern refers to the pattern in the total bits of data sent (TBS) from a cell tower to all cell phones interacting with that tower over time. Here, a cell phone's traffic pattern refers to the pattern in the TBS from a cell tower to the specific cell phone, while the two interact over time. Both traffic patterns are represented as time series, which consists of the TBS from a tower, indexed over units of time (timesteps). For both patterns, we use a GRU model to predict the amount of data, i, that will be sent j timesteps in the future, based on the value of i in each of the k most recent timesteps. The specific meanings and units of i; j; k are detailed in Chapter 5. The project completes two objectives. First two trafic pattern datasets are collected and processed into time series form, one of a cell tower's pattern over 30 days and one of a cell phone's pattern in the context of watching a YouTube video. Second, seven data featurization approaches are implemented and the effects of each on model prediction error, measured by RMSE, are assessed; approaches are also compared to a baseline approach and the GRU is compared to naive prediction models that do not use machine learning. Regarding cell tower data, the GRU model achieves low prediction error (11.63% RMSE for baseline featurization) but it relies on high autocorrelation in the time series and is only marginally better than the naive models. Only one featurization approach marginally improves GRU performance on cell tower data (11.54% RMSE). Regarding the cell phone data, the GRU model achieves a low prediction error (16.50% RMSE for baseline featurization) even though the time series has low autocorrelation. Featurization approaches, that increase autocorrelation, reduce error (3.75% RMSE), although the GRU model resembles naive models in these cases. This work lays the foundation for improving congestion control in cellular networks and the experience of network users.