Research On Prediction For Quality Of Transmission And Optimization For Resource Allocation In Optical Networks Based On Machine Learning

With the rapid growth of mobile traffic and the rise of diversified services in the 5G era,the development of optical networks faces many challenges from the transmission control level to resource optimization.In a transmission control domain,optical networks need to be dynamically configured to support the service in an on-demand way.In terms of resource optimization,mobile fronthaul optical networks require a flexible functional split scheme to achieve a service-oriented baseband function deployment.Finally,due to the mobility and variability of user traffic,the baseband function chain needs to be adjusted frequently to maintain resource utilization.The facts that optical networks tend to be dynamic,the deployment of baseband functions becomes more complicated,and the environment is uncontrollable make the manual experience-dominated mechanism unsustainable for resource allocation,and it requires an automated and intelligent operation and maintenance mechanism.Machine learning,learning from data and continuously improving the strategy by imitating human learning,enables optical networks to establish the intelligent operation and maintenance mechanism.Therefore,machine learning based accurate quality of transmission(QoT)prediction model and the efficient strategy of resource optimization are investigated from the perspectives of "data mining","decision reasoning" and "user perception",which leads to intelligent optical networks gradually.The main contributions and innovations in the research are as follows:(1)An artificial neural networks(ANN)based multi-channel QoT prediction model is proposed.To predict the QoT of new establishing channels and existed channels in optical network reconfigurations,an ANN based multi-channel Q-factor regression model is proposed.The proposed model considers the impacts of the configuration status,link informations,real-time parameters of optical devices,spectral information and other factors on the channel's QoT.Based on the collected data in real transmission scenarios,an ANN based multi-channel Q-factor prediction model is explored in the paper.Compared with the traditional physical analytical model for QoT estimation,the model proposed in this paper can predict the QoT quickly and accurately.Experimental results show that the ANN based Q-factor regression model can precisely predict the Q-factors of all the channels at the same time,the mean absolute error(MAE)on the test is less than 0.1 dB.This comprehensive and accurate prediction enables dynamic optical networks to have a global view of all the channels,which benefits the efficiency and stability of network reconfiguration.(2)A deep reinforcement learning(DRL)-based policy for the deployment of based baseband function(BBF)and routing is proposed.To address the great pressure of bandwidth for 5G fronthaul optical networks,the policy of BBF deployment and routing based on DRL is proposed in the paper.Based on the data from the interaction between the agent and the network environment,an ANN based Q-function is fitted on the collected data to output the values of actions under the different policies precisely.After the fitting of Q-function,the proper policy is selected through the action's value by ANN.The simulation results show that the proposed DRL-based policy achieves a superior performance to the traditional heuristic algorithm,First-Fit algorithm both in centralized-radio access networks(C-RAN)and next generation-radio access networks(NG-RAN),and the performance is also very close to the optimal solution by integer linear programming(ILP).Unlike the fact that the ILP model is only applicable to offline scenarios,the DRL-based algorithm has also achieved the satisfied performance in dynamic online scenarios.This provides a new technical route for resource optimization in dynamic scenarios.(3)The strategy for BBF deployment and migration is proposed based on traffic prediction.To address the decrease in resource utilizations and the service interruptions by the traffic change in mobile networks,the strategy of BBF deployment and migration based on time-series traffic prediction is proposed,which aims to balance the service interruptions by BBF migration and the resource efficiency.Different from the traditional optimization strategies that only consider the inherent resources of the network,the proposed strategy in this paper builds a long and short memory network(LSTM)based model to predict the time-series traffic by collecting historical traffic data from operators.Once the model works,a joint ILP model based on the predicted traffic and network resources is proposed to generate the strategy for BBF deployment.Simulation results show that the proposed strategy can reduce the number of user migrations while maintaining high resource utilizations.The mechanism for resource allocation based on the traffic prediction in advance can procactively address the influence of the mobile traffic on network.