基于多时间粒度时空图网络的蜂窝网络流量预测

Cellular Network Traffic Prediction Based on Spatio-temporal Graph Network with Multiple Temporal Granularity

蜂窝网络流量预测对于运营商提高网络服务质量、降低能耗、优化资源配置具有重要意义。针对当前蜂窝网络流量预测方法无法同时提取多时间粒度序列特征和空间特征的问题,提出一种基于多时间粒度时空图神经网络的蜂窝网络流量预测方法。该方法首先将基站历史数据建模为多个时间粒度的时序数据,并使用一维卷积网络提取每个序列的特征,然后使用图注意力网络对多时间粒度的特征进行聚合得到单一基站的数据特征,最后将多个基站的特征进行空间聚合,并使用全连接层将每个基站聚合后的特征映射为预测结果。实验选择公开数据集Telecom Italia验证该方法的有效性,使用RMSE和R 2作为预测结果的评价指标,与当前已有方法相比该方法可取得最好的预测结果。论文最后分析了不同时间粒度序列对最终预测结果的影响,结果表明时间粒度位于40分钟至1.5小时之间的序列对提高模型预测效果的贡献最大。

Cellular network traffic prediction is of great significance for operators to improve network service quality,reduce energy consumption,and optimize resource allocation.A cellular network traffic forecasting approach based on a multi-time granularity spatio-temporal graph neural network is proposed to address the problem that current cellular network traffic forecasting methods cannot extract multi-time granularity sequence features and spatial features effectively.The historical traffic data of the base station is modeled as time series of multiple time granularities,and the one-dimensional convolutional network is applied to extract the features of each sequence.Then the graph attention network is employed to aggregate the features of multi-time granularities to obtain the embedding of a single base station.Finally,the embeddings of multiple base stations are spatially aggregated,and the final prediction result for each base station is obtained via a fully connected network.The public dataset Telecom Italia is used to verify the effectiveness of the proposed method,and RMSE and R 2 are used as evaluation indicators for the prediction results.The proposed method can achieve the best prediction results compared with the current existing methods.Finally we analyze the influence of different time granularity sequences on the final prediction results.The results show that sequences with time granularity between 40 minutes and 1.5 hours make the greatest contribution to improving the model prediction effect.