Utilizing an Ensemble STL Decomposition and GRU Model for Base Station Traffic Forecasting

Owing to the growing sophistication of rapidly developing wireless communications networks and systems, corresponding network traffic volume has similarly scaled in both proportion and complexity, posing a challenge in meeting user demand. Base station traffic forecasting is at the frontline of research currently as a response to meet the needs of future users. We propose an ensemble model that combines time series decomposition via Seasonal and Trend Decomposition using Loess (STL), individually forecast decomposed components with gated recurrent unit (GRU) neural network model, then recombine forecasts. The rationale behind the hybrid model is that the decomposition reduces the effect of noise and outliers in the time series data, thereby enabling improved results when forecasting traffic data as compared standalone statistical or machine learning techniques. Our proposed scheme shows performance gains in model forecast accuracy.