Dynamic Learning Framework for Smooth-Aided Machine-Learning-Based Backbone Traffic Forecasts

Sensors

Mosab Hamdan¹  Monia Hamdi²  Habib Hamam³  Mohamed Khalafalla Hassan⁴  N. Effiyana Ghazali⁴  Sharifah Hafizah Syed Ariffin⁴  Suleman Khan⁵  Mutaz Hamad⁶ 

[1] Department of Computer Science, University of São Paulo, São Paulo 05508-090, Brazil;Department of Information Technology, College of Computer and Information Sciences, Princess Nourah Bint Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia;Faculty of Engineering, Uni de Moncton, Moncton, NB E1A3E9, Canada;School of Electrical Engineering, University Technology Malaysia, Skudai, Johor 81310, Malaysia;School of Psychology and Computer Science, University of Central Lancashire, Preston PR1 2HE, UK;School of Telecommunication Engineering, Future University, Khartoum 10553, Sudan;

关键词: traffic forecast;    slice;    local smoothing;    LSTM;    dynamic learning;   

DOI  :  10.3390/s22093592

来源: DOAJ

【 摘 要 】

Recently, there has been an increasing need for new applications and services such as big data, blockchains, vehicle-to-everything (V2X), the Internet of things, 5G, and beyond. Therefore, to maintain quality of service (QoS), accurate network resource planning and forecasting are essential steps for resource allocation. This study proposes a reliable hybrid dynamic bandwidth slice forecasting framework that combines the long short-term memory (LSTM) neural network and local smoothing methods to improve the network forecasting model. Moreover, the proposed framework can dynamically react to all the changes occurring in the data series. Backbone traffic was used to validate the proposed method. As a result, the forecasting accuracy improved significantly with the proposed framework and with minimal data loss from the smoothing process. The results showed that the hybrid moving average LSTM (MLSTM) achieved the most remarkable improvement in the training and testing forecasts, with 28% and 24% for long-term evolution (LTE) time series and with 35% and 32% for the multiprotocol label switching (MPLS) time series, respectively, while robust locally weighted scatter plot smoothing and LSTM (RLWLSTM) achieved the most significant improvement for upstream traffic with 45%; moreover, the dynamic learning framework achieved improvement percentages that can reach up to 100%.

【 授权许可】

Unknown