基于深度学习的人群活动流量时空预测模型

传统的时空预测方法缺乏对复杂时空非线性关系的描述,且难以顾及空间多尺度特征对于预测结果的影响。针对这一问题,本文提出了一种融合空间多尺度特征的时空网络模型(MST-Net),将流量预测的回归问题转换为具有时空特性的判别模型。首先,通过并联卷积提取空间多尺度特征;然后,通过引入注意力机制的门控循环单元提取时间特征;最后,利用全连接层得到预测结果。本文将该模型用于人群活动流量的预测,分别在两组真实的社交媒体签到数据集上进行试验。试验结果表明:本文采用的卷积层连接方式和特征融合方法,相比于单层卷积层提取空间特征、其他连接方式和融合方法以及传统的时空预测模型,在均方根误差(RMSE)和平均绝对百分比误差(MAPE)两个预测结果评价指标上均有不同程度的提高,说明本文方法具有较高的预测精度,能够较好地拟合时空问题的非线性关系,实现人群活动流量的预测。

A spatio-temporal network for human activity prediction based on deep learning

The traditional spatio-temporal prediction methods could hardly model the complex nonlinear relationship of spatio-temporal phenomenons,thus they lack the ability to consider the influence of spatial multi-scale characteristics into the prediction results.In order to overcome this deficiency,a novel model of space-time network(MST-Net)is proposed in this paper,which transforms the regression problem of volume prediction into a discriminant model with time-space characteristics.The spatial and temporal characteristics of spatio-temporal data are extracted by multi-scale parallel convolution and gate recurrent unit respectively.Thus the extracted features are fused with the attention mechanism introduced to capture the long-term features.Finally,the prediction results can be obtained by using the full connection layers.In order to prove the reliability and validity of the model,the model is tested on two challenging social media sign-in datasets.The results indicate that the proposed model outperformed other algorithms in two prediction results evaluation indexes,namely the root mean square errors(RMSE)and mean absolute percentage errors(MAPE),which illustrate that the proposed method could achieve higher prediction accuracy and could better fit the nonlinear relationship of the space-time problem.The proposed model is suitable to predict the flow of human activities.