Research On Machine Learning Based Routing And Spectrum Allocation And Fragmentation Mechanisms In Elastic Optical Networks

Traditional Wavelength Division Multiplexing?WDM?optical network uses fixed-size wavelengths as the minimum granularity of resource allocation.It is necessary to add protection bandwidth between different wavelength bands,which is difficult to adapt to dynamic bandwidth requirements and resulting in poor spectrum resources utilization.Elastic Optical Networks?EON?introduces the Optical Orthogonal Frequency Division Multiplexing?O-OFDM?and supports variable granularity resource allocation,which has the advantages of higher bandwidth resource utilization and more flexible bandwidth allocation.EON has become one of promising solution for future optical networks.Aiming at the problem of routing and spectrum resource allocation in EON,this thesis introduces an improved RSA algorithm based on machine learning and traffic prediction,and studies it through theoretical analysis and numerical simulation.Firstly,the thesis briefly describes some basic principles,architecture and key technologies of the EON,then analyses the existing routing and spectrum allocation algorithms,discusses and compares the methods of predicting time series and the principles and formulas of common models,respectively.In view of the network congestion and spectrum fragmentation problems that may be caused by burst traffic in the elastic optical network,predicting the service traffic in the link based on machine learning is proposed.After theoretical research,the thesis proposes a method to predict the traffic flow of EONs.The concept of link congestion is defined and Recurrent Neural Network?RNN?as well as simulation data are used to predict the congestion of each link in three common network topologies.In order to verify the proposed method,this thesis designed and constructed the training dataset and test dataset used in the model for the National Science Foundation Network?NSFNET?topology,and used the training to predict traffic.The simulation results show that the short-term prediction accuracy of the prediction method proposed in this paper can reach 0.95.Aiming at the RSA problem in elastic optical networks,the paper proposes prediction-based routing and spectrum allocation algorithms.The improved algorithm searches the routing table in advance to determine K candidate paths before new services arrive.When a new service arrives,first select the shortest path from the candidate paths and determine whether the number of candidate paths is zero.After that,it calculates and determines whether the searched shortest path congestion degree?is greater than 1and time t=₀.Next step is determineing whether there is an available frequency slot on the path and perform spectrum allocation.Theoretical analysis and simulation results show that,compared with the traditional algorithm,the service blocking rate in NSFNET,China Education and Research Network?CERNET?,and United States Network?USNET?can be improved by 7%,12%,and8%when the network load is 600Erl,respectively.