An intelligent traffic prediction framework for 5G network using SDN and fusion learning

Traffic prediction and analysis is an important part of traffic engineering in Software Defined Networking (SDN). Effective forecasting of network traffic needs the composition of capturing the dependency features. SDN is embedded in 5G networks which decouples control plane from the data plane for network programmability. Flow based SDN architecture provides precise prediction of network traffic with fine granularity of data plane network statistics. Deep learning models like Gated Recurrent Unit (GRU) enables the time series forecasting with high performance. The major problem with usage of deep learning models is its communication overhead with the convergence of the model. To provide a communication efficient and intelligent traffic prediction framework, fusion learning is used between the data plane and control plane of the SDN environment. Fusion learning provides the prediction model with the exchange of model parameters of the SDN client models and its data distribution with single communication. The global topology manager of the SDN controller with the deep neural prediction model enhances the forecasting of the network traffic. Time series GRU captures only the temporal dependency. To handle dynamics of the network traffic and efficient capture of traffic pattern, spatial dependency must be captured. The diffusion convolution operation embedded in the GRU capture both spatial and temporal dependency of the features in encoder-decoder architecture. The stochastic gradient based scheduled sampling improves the performance of the prediction model with the optimal decay rate. The proposed framework is tested with the simulated data on Abilene network topology with RYU SDN controller. The experimental results exhibit improved accuracy in both local and global model of 87%—94%. Further, hyperparameter tuning is done for precise forecasting of network traffic with the minimal prediction error of 7.98% for one hour horizon. The proposed prediction model observes improvement over the base line prediction models.