Accurate Load Prediction Algorithms Assisted with Machine Learning for Network Traffic

With the increasingly higher demand on radio access networks, problems such as serious energy consumption and network load imbalance have aroused, catching more attention from the network operators. To solve these problems, load prediction has been considered as one possible solution. This paper investigates two load prediction models to predict the traffic load in cells, in order to improve the energy saving strategies. One of the proposed algorithm is a linear ensemble model which is composed of three sub-models, to predict the traffic load in terms of time, space and historical pattern respectively. This model is named as the ensemble model. Different methods such as time series analysis, linear regression and regression tree are applied to the sub-models, after which adjusted weights are calculated and allocated to each of the three sub-models to create the ensemble model. The second model investigated in this paper is convolutional neural network (CNN) based, which utilizes the residual learning network (ResNet) structure to train the collected data and at the same time improves the calculation efficiency. This model is thus referred as ResNet model in this paper. The collected data of traffic load in Milan area is put into training and verification by the proposed models. The performance of proposed methods is evaluated. The result shows that ensemble model provides a higher prediction accuracy than the other base line models and the ResNet model improves the efficiency of calculation.