Short-term traffic forecasting for a smart satellite communications system

Abstract

Satellite communications systems are undergoing a modernization to efficient capacity allocation from a traditional "bent pipe" or static allocation. One challenge to address with a more precise usage of satellite resources is the change in user terminal traffic during a complete cycle of the system: collecting data, generating a constellation setting solution, transmitting the new solution to each satellite and executing changes to the satellite's parameters. As the system's cycle time grows the user's desired data rate changes causing an optimized solution based on an erroneous traffic model. This thesis proposes a comparison of single user models using a gradient boosting algorithm, and a multi-user model using Long- Short Term Memory neural networks (LSTM) or Gated Recurrent Unit neural networks (GRU) to forecast terminal traffic. Each algorithm was tuned using a two-stage design of experiments process consisting of a fractional screening design to identify impactful hyper-parameters and a central composite design to find optimal model settings. During a holdout period, the mean absolute percentage error using a 15-minute lag was 10.7% with a standard deviation of 2.6% over a month of forecasting. Networks using a GRU layer and tuned with Random Search had the best average performance with an error of 9.6% and standard deviation of 2.6%, outperforming the best found XG Boost models with an error of 9.9% and standard deviation of 3.5%.