Multi-scale 4D localized spatio-temporal graph convolutional networks for spatio-temporal sequences forecasting in aluminum electrolysis

Abstract

Spatio-temporal sequences forecasting fulfills a vital role in the intelligent advancement of aluminum electrolysis production process. The localized spatio-temporal correlations contained in spatio-temporal sequences, caused by the dynamicity of regional working conditions, have complex and diverse (multi-scale) characteristics. The existing deep learning-based prediction methods are difficult to capture the multi-scale localized spatio-temporal correlations, and the adverse effects of industrial noise on spatio-temporal correlation acquisition have not been considered. In this article, we propose the multi-scale 4D localized spatio-temporal graph convolutional networks (Ms-4D-LStGCN) to address the above issues. Concretely, we propose a data-driven accurate similarity search method and fuse the prior knowledge to construct the spatio-temporal graph. Then,a novel 4D localized spatio-temporal graph convolution module is proposed to capture the complex localized spatio-temporal correlations. Finally, the multi-scale 4D localized spatio-temporal graph convolution framework is designed to obtain the multi-scale and multi-depth localized spatio-temporal correlation features. Illustrative examples on 16 real-world industrial aluminum electrolysis datasets attest that our method has superior prediction performance compared with state-of-the-art methods.