A novel spatial-temporal graph convolution network based on temporal embedding graph structure learning for multivariate time series prediction

The prediction of multivariate time series is a pivotal research area in data mining, offering extensive practical applications in many real-world scenarios, including transportation, finance, energy systems, the Internet of Things. Accurately predicting multivariate time series is challenging due to the complex temporal and spatial dependencies among variables. To tackle this challenge, this study proposed a deep learning model utilizing graph neural networks for predicting multivariate time series. Specifically, a multivariate time series is modeled as a graph, with nodes representing variables, edges indicating their interdependencies, and the time series data serving as node attributes. We leverage the temporal convolutional network to construct a graph structure learning module that captures the underlying dependencies between variables through the learned adjacency matrix. The prediction model was built by integrating Long Short-Term Memory networks and graph neural networks, enabling the simultaneous capture of temporal and spatial dependencies in multivariate time series data. Additionally, to mitigate the issue of over-smoothing in graph neural networks, we incorporated the Laplacian sharpening technique into our model. The proposed method is generalizable for handling multivariate time series data, as it does not require a pre-defined adjacency matrix among variables. We empirically evaluated the performance of our method through extensive experiments conducted on six real-world datasets, the experimental results indicated that our method could effectively improve the accuracy of multivariate time series prediction.