Spatio-temporal graph convolutional networks driven by data-physical fusion for parameter prediction of natural gas dehydration system

Abstract

Triethylene glycol (TEG) dehydration unit is a piece of essential device for removing moisture from raw natural gas during natural gas production. However, the existing station equipment management systems are mostly collection-oriented with little analysis, lack the effective methods of parameter prediction and fault warning, and the strong coupling between the monitoring parameters is a problem should be study. To solve these problems, this paper analyzes the time dependence and spatial correlation of these parameters. Also, a spatio-temporal graph convolutional networks prediction model driven by data-physical fusion (SG-STGCN) is proposed for constructing the graph structure. Firstly, the signed directed graph (SDG) model is established based on the physical process, and the weight of each edge is obtained by using the grey relational analysis (GRA). Secondly, by stacking spatio-temporal convolutional modules, the temporal and spatial dependencies over a long range of time are captured to realize multivariate parameter prediction. Then, the real-time monitoring data of a dehydration station are used for analysis. The experimental results showed that the proposed method can achieves the best predict result compared with other methods, and can be used in the fault early warning to maintain high reliability of equipment. Finally, the SG-STGCN has been integrated and tested successfully on the real-time monitoring platform of a dehydration unit.