【摘要】 通信基站负载预测在移动通信网络管理与优化中具有重要作用，主要挑战是准确建模时空依赖特征。本文提出基于时空层次图神经网络（spatial-temporalhierarchical graph neural network, ST-HGNN）的通信基站负载预测模型，首先设计顾及空间邻接性和时序相关性的时空节点聚类算法，构建基站与区域层次图；继而利用时空模块提取局部时空特征，并构建基于注意力机制的特征融合模块，充分识别区域内和跨区域的层次交互，捕获非局部时空特征；最后引入日期类型等外部特征，通过全连接层输出基站负载预测结果。实验结果表明，在长沙市两个行政区（各435/399个基站）2019—2020年两周的负载数据上，本模型相较最先进方法在RMSE指标提升3.92%以上，MAE指标提升2.44%以上，消融实验证实了时空节点聚类算法和基于注意力机制的特征融合模块的有效性。更多还原

【Abstract】 Mobile traffic forecasting plays a crucial role in the management and optimization of mobile networks,and its main challenge is accurately modeling spatial-temporal dependencies. This paper proposes the Spatial-Temporal Hierarchical Graph Neural Network(ST-HGNN) approach. First, a spatial-temporal node clustering method is designed to construct the hierarchical structure. Second, a convolution encoder is employed to extract the local spatial-temporal features from a hierarchical perspective. Furthermore, a novel attention-based feature fusion module is built to capture the non-local spatial-temporal features by identifying both the intra-regional and cross-regional feature impacts on stations. Finally, stacked fully connected layers with external feature(date type) is introduced to produce the final prediction. Extensive experiments on two real-world datasets demonstrate that the performance of our model outperforms the state-of-the-art methods, while the ablation study also verifies the effectiveness of the spatial-temporal node clustering and the attentionbased feature fusion module.更多还原