Unlicensed Spectrum Forecasting: An Interference Umbrella Based on Channel Analysis and Machine Learning

Abstract

In this work a novel framework for predicting future interference levels for IEEE 802.11 networks is developed and experimentally evaluated. At the heart of the framework lies a modelling mechanism which is able to estimate and determine in real-time, the over-the-air performance that each network user will receive over an IEEE 802.11 link, when considering and combining multiple wireless metrics, without requiring a network association (Wi-Fi AP-STA) to be performed. On top of the solution, a Machine Learning approach is integrated, in order to project the real-time predictions to long-term predictions in the future (2-hour interval). Additionally, the framework applies a self-correcting mechanism for the predictions, by extracting short-term predictions and accurate throughput calculations, when the current channel conditions largely differ from the long-term predictions. The proposed framework covers comprehensively the cases of interference created by either 802.11 or non 802.11 devices, which may occur at the target or at any overlapping wireless channel. Finally, extensive testbed experimentation proves the framework's proper functionality and accuracy, under the cases of both Indoor (controlled interference) and Outdoor (uncontrolled massive interference) environments.