A Novel Multiscale Transformer Network Framework for Natural Gas Consumption Forecasting

Yanyun Pu, Chengyuan Zhu, Kaixiang Yang, Zhuoling Lü, Qinmin Yang

Abstract: Accurate and timely natural gas consumption forecasts are essential for energy policy formulation, natural gas scheduling, and pipeline network design. However, it remains a significant challenge because natural gas consumption is highly nonlinear and irregular with complex cycles. In this article, we propose a new spatial-temporal multiscale transformer network framework that exploits dynamic spatial dependence among users and temporal support of historical multivariate data to improve the accuracy of short-term natural gas consumption forecasting. A novel graph neural network model is developed to capture the spatial dependencies relationships among users by considering the fixed and dynamic connectivity. Compared with other approaches, we validate the effectiveness of the proposed model and its ability to capture fine-grained and spatial-temporal dependencies on a real dataset.