基于时空特征的无线网络流量预测方法

Prediction Method of Wireless Network Traffic Based on Spatiotemporal Features

无线网络流量分布具有空间上和时间上的特征,针对传统预测方法对流量分布空间特征的利用不足问题,提出三维卷积神经网络(3D-CNN)和长短期记忆网络(LSTM)相结合的无线网络流量预测模型。首先通过3D-CNN挖掘流量数据的局部时空关联性,并利用空间注意力机制完善全局空间关联的提取;然后使用LSTM模型对抽象时空特征进行训练,并加入了注意力机制缓解循环神经网络的遗忘现象带来的信息损耗。运用此方法对"意大利电信大数据挑战赛"的公开数据集进行训练,其均方根误差(RMSE)和平均绝对误差(MAE)分别降至5.17和3.32,明显优于其他对比预测模型。

The distribution of wireless network traffic has spatial and temporal characteristics. Aiming at the insufficient utilization of the spatial characteristics of traffic distribution by traditional prediction methods, a wireless network traffic prediction model combining three-dimensional convolutional neural network(3D-CNN) and long short-term memory network(LSTM) is proposed. First, 3D-CNN is used to mine the local spatiotemporal correlation of traffic data, and the spatial attention mechanism is used to improve the extraction of global spatial correlation;then the LSTM model is used to train the abstract spatiotemporal features, and the attention mechanism is added to alleviate the forgetting of the recurrent neural network. information loss caused by the phenomenon. Using this method to train the public dataset of "Telecom Italia Big Data Challenge", its root mean square error(RMSE) and mean absolute error(MAE) are reduced to 5.17 and 3.32, respectively, which is significantly better than other comparative prediction models.