Forecasting Wireless Demand with Extreme Values using Feature Embedding in Gaussian Processes

TL;DR

This paper introduces a feature embedding kernel for Gaussian Processes to improve the prediction of extreme wireless traffic demand spikes and troughs, providing better accuracy and probabilistic uncertainty quantification compared to existing models.

Contribution

The paper proposes a novel FE kernel for GPs that enhances extreme value forecasting and offers a flexible trade-off between overall and peak-trough accuracy.

Findings

32% reduction in short-term extreme value prediction error vs. S-ARIMA

21% reduction in long-term average prediction error vs. S-ARIMA

Provides probabilistic forecast uncertainty for risk assessment

Abstract

Wireless traffic prediction is a fundamental enabler to proactive network optimisation in beyond 5G. Forecasting extreme demand spikes and troughs due to traffic mobility is essential to avoiding outages and improving energy efficiency. Current state-of-the-art deep learning forecasting methods predominantly focus on overall forecast performance and do not offer probabilistic uncertainty quantification (UQ). Whilst Gaussian Process (GP) models have UQ capability, it is not able to predict extreme values very well. Here, we design a feature embedding (FE) kernel for a GP model to forecast traffic demand with extreme values. Using real 4G base station data, we compare our FE-GP performance against both conventional naive GPs, ARIMA models, as well as demonstrate the UQ output. For short-term extreme value prediction, we demonstrated a 32\% reduction vs. S-ARIMA and 17\% reduction vs. Naive-GP. For long-term average value prediction, we demonstrated a 21\% reduction vs. S-ARIMA and 12\% reduction vs. Naive-GP. The FE kernel also enabled us to create a flexible trade-off between overall forecast accuracy against peak-trough accuracy. The advantage over neural network (e.g. CNN, LSTM) is that the probabilistic forecast uncertainty can inform us of the risk of predictions, as well as the full posterior distribution of the forecast.