Short term prediction of wireless traffic based on tensor decomposition and recurrent neural network

Abstract

This paper proposes a wireless network traffic prediction model based on Bayesian Gaussian tensor decomposition and recurrent neural network with rectified linear unit (BGCP-RNN-ReLU model), which can effectively predict the changes in the upstream and downstream network traffic in a short period of time in the future. The research is divided into two parts: (i) The missing observations are imputed by an algorithm based on Bayesian Gaussian tensor decomposition. (ii) The recurrent neural network is used to forecast the true observations only rather than both true and estimated observations. The results show that, compared with other combined models of missing data imputation and neural networks, the BGCP-RNN-ReLU model proposed in this paper has the smallest prediction error for both the upstream and downstream traffic. The new model achieves better forecasting precision, and thus can help to regulate the load of communication station to reduce resource consumption.

Highlights

The problem of forecasting wireless network traffic with missing values is divided in two stages to handle.

A newly propose d method can more efficiently impute missing values in wireless network traffic data.

Simple recurrent neural network obtains better prediction performance than other complex networks.