针对设备端口链路的LSTM网络流量预测与链路拥塞方案

针对设备端口链路流量，提出两种基于长短期记忆网络的预测模型。第一种针对在大时间粒度下平稳变化的流量；第二种则针对在小时间粒度下波动剧烈的非平稳流量。通过选用不同的数据划分方式与模型训练方法，构建两种具有不同网络结构的流量预测模型。实验结果表明，前者在处理平稳变化的流量时能够达到极高的预测精度，后者在处理非平稳流量时具有明显优于SVR模型、BP神经网络模型的预测效果。在第二种预测模型的基础上，提出了参数可调的链路拥塞预警方案，实验证明该方案具有一定的可行性。

LSTM network traffic prediction and link congestion warning scheme for single port and single link

To predict the traffic at single port and single link,two network traffic prediction models based on long short-term memory neural network were proposed.The first model is for the traffic which changes smoothly at large time granularity.The second model is for the nonstationary traffic which fluctuates violently at small time granularity.By selecting different methods of splitting data and training models,two traffic prediction models with different neural network structures were constructed.The experimental results show that the former can achieve a very high accuracy when predicting smoothly changed traffic,the latter has a significantly better prediction effect than the support vector regression model and the back propagation neural network model when dealing with nonstationary traffic.Based on the second model,a link congestion warning scheme with variable parameters was proposed.The scheme is proved to be practicable by experiments.