基于全注意力机制的多步网络流量预测

Predicting resource allocation can effectively utilize the remaining resources of wireless networks to serve non-real-time services. One of the key issues among them is the prediction of remaining resources, which can be transformed into the problem of real-time traffic prediction. In this paper, the full attention mechanism proposed for natural language processing is introduced into the time series prediction problem to predict the traffic at the second level within a future minute-level time window. Through training and testing on the measured traffic data set recorded every second, it is compared with other methods based on recurrent neural networks and linear and nonlinear prediction models in terms of complexity (measured by training and testing time), prediction accuracy (measured by mean relative percentage error), and statistical characteristics of prediction errors (measured by the mean and standard deviation of prediction errors). The research results show that compared with the recurrent neural network without an attention mechanism, the designed method based on the full attention mechanism has a lower computational complexity, and due to the cumulative error of multi-step prediction, the gain in prediction accuracy is not obvious.

预测资源分配能有效利用无线网络的剩余资源服务非实时业务,其中的关键问题之一是剩余资源的预测,可转化为实时业务流量预测问题。本文把面向自然语言处理提出的全注意力机制引入到时间序列预测问题中,预测未来分钟级时间窗内秒级的流量,通过在每秒记录的实测流量数据集上进行训练和测试,与其他基于循环神经网络和线性、非线性预测模型的方法在复杂度(由训练和测试时间衡量) 、预测精度(由平均相对百分比误差衡量)和预测误差统计特性(由预测误差的均值和标准差衡量)等方面进行比较。研究结果表明,与无注意力机制的循环神经网络相比,所设计的基于全注意力机制的方法计算复杂度低,由于多步预测的累积误差,在预测精度方面增益不明显。