基于动态稀疏注意力的地铁客流预测模型

Metro Passenger Flow Prediction Model Based on Dynamic Sparse Attention

地铁客流预测是随时间演变的多维时间序列数据,不同序列之间存在复杂的动态相互依赖关系。为挖掘多种监测指标之间存在的内在复杂关系,提出动态稀疏注意力模型:利用全局变量注意力自动选择相关驱动序列,增强模型预测的判别性;根据局部紧密相关和全局稀疏相关的先验知识,对历史时间步和相关变量分别卷积和稀疏卷积,提取局部时间和局部变量特征;设计了稀疏注意力对相关时间步加权和变量加权,提高预测表现。结果表明,与其他常用客流预测模型相比,动态稀疏注意力模型能高度准确地预测客流。

Metro passenger flow prediction is multi-dimensional time series data that changes as time passes,and complex dynamic interdependent relationship exists among various series.To explore the internal complex relationship among various monitoring indicators,dynamic sparse attention model(DSANNs)is proposed,which uses global variable attention to automatically select relevant driver sequence to enhance the discriminability of model prediction.According to the prior knowledge of local close correlation and global sparse correlation,the convolution and sparse convolution of historical time steps and related variables are conducted to extract the characteristics of local time and local variables,and the weight of related time steps and variables by sparse attention is designed to improve prediction performance.The experiment results show that,compared with other commonly used passenger flow prediction model,DSANNs can carry out the prediction with high level of accuracy.