Time Series Prediction in Software – Defined Network Using Deep Learning

Abstract

This paper presents the results of the study on time series prediction possibilities for acquired traffic bandwidth data in a software-defined network (SDN), by using a deep learning (DL) system in form of a recurrent neural network (RNN). Researching the tuning of RNN hyperparameters the paper examined time window for time series past data, batch size of analysed data, as well as time window for prediction of time series values in the future, number of epochs and lastly, steps per epoch for an RNN training process. Further, for an SDN simulation a Mininet emulator and an OpenDaylight SDN controller was used. An RNN was implemented in Keras application programming interface for Google TensorFlow machine learning (ML) platform. It was confirmed that RNNs present a significant alternative to traditional stochastic models for time series prediction.