基于时序特征建模的往返时延预测方法

Round-trip time prediction method based on time series feature modeling

为提升往返时延预测的准确性与实时性,在深入挖掘和分析影响其准确度各种因素的基础上,针对往返时延变化的短期随机性与长期平稳性,提出一种基于时序特征建模的往返时延预测方法GCA-RTT。通过构建门控卷积与自注意力机制相融合的时延历史数据局部特征与长期依赖关系学习模型,实现更为精确、高效的往返时延预测。实验结果表明,GCA-RTT可以有效捕捉基于时间序列的往返时延变化特征,与其它神经网络预测方法比较,预测准确性明显提高且预测时间缩短。

To improve the accuracy and real-time performance of round-rip time prediction,on the basis of in-depth mining and analysis of various factors affecting its accuracy,the round-trip time prediction method based on time series feature modeling GCA-RTT was proposed for the short-term randomness and long-term stability of round-trip time variation.A more accurate and efficient round-trip time prediction was achieved by constructing a local feature and long-term dependence learning model of historical time delay data integrated with gated convolution and self-attention mechanism.Experimental results show that GCA-RTT can effectively capture the round-trip time variation feature based on time series,and the prediction accuracy is significantly improved and the prediction time is shortened compared with that of other neural network prediction methods.