Spatiotemporal Causal Discovery Graph Convolutional Networks for Multivariate Time Series Forecasting of Industrial Process

Traditional neural network methods cannot make full use of the potential relationships between multiple variables in industrial processes due to their nonlinearity and coupling complex relationship, then multivariate time series forecasting is challenging. In this paper, a novel graph neural network model for multivariate time series forecasting is proposed based on spatiotemporal causal discovery. First, a spatiotemporal causal discovery algorithm is designed to extract the potential relationships between multiple temporal variables, then the multivariate time series spatiotemporal causal graph is constructed. Next, the graph convolution network module and the gated temporal convolution network module are integrated to extract the potential causal relationship, short-term dependence and long-term dependence of multiple variables. Finally, Experiments on a real industrial dataset show that our proposed method has higher prediction accuracy and broad application prospects in industrial processes and other complex systems of multivariate time series.