基于时空图卷积网络的多变量时间序列预测方法

为了扩大时空图卷积网络的预测范围，将它应用在关联关系未知场景下的多变量时间序列预测问题，提出一种附加图学习层的时空图卷积网络预测方法（GLB-STGCN）。图学习层借助余弦相似度从时间序列中学习图邻接矩阵，通过图卷积网络捕捉多变量之间的相互影响，最后通过多核时间卷积网络捕捉时间序列的周期性特征，实现对多变量的精准预测。为验证GLB-STGCN的有效性，使用天文、电力、交通和经济四个领域的公共数据集和一个工业场景生产数据集进行预测实验，结果表明GLB-STGCN优于对比方法，在天文数据集上的表现尤为出色，预测误差分别降低了6.02%、8.01%、6.72%和5.31%。实验结果证明GLB-STGCN适用范围更广，预测效果更好，尤其适合自然周期明显的时间序列预测问题。

Multivariate time series forecasting with spatio-temporal graph convolutional networks

In order to expand the prediction range of spatio-temporal graph convolutional networks and apply them to the multivariate time series prediction problems in the scenario of unknown correlation, this paper proposed a graph learning based spatiotemporal graph convolutional networks(GLB-STGCN). The graph learning layer learned the graph adjacency matrix from the time series with the help of cosine similarity, then the graph convolution networks captured the interaction between multi-variables, and finally the multi-kernel time convolution networks captured the periodic characteristics of the time series for precise prediction. To verify the effectiveness of GLB-STGCN, this paper used 4 public datasets from astronomy, electricity, transportation and economy and 1 industrial production dataset for the prediction experiments. The results show that GLB-STGCN outperforms the comparison methods, especially on astronomical datasets, with prediction errors reduced by 6.02%, 8.01%, 6.72%, and 5.31%, respectively. The experimental results prove that GLB-STGCN has a wider application range and better prediction effect, especially for time series prediction problems with obvious natural cycles.