A graph neural network (GNN) method for assigning gas calorific values to natural gas pipeline networks

Abstract

Calorific value metering offers unique advantages in the natural gas industry. Based on a graph neural network (GNN), this study proposes a method for metering the gas calorific value in natural gas transportation. Graph theory is introduced to consider the effects of physical topology and node properties in natural gas pipeline networks. Inductive inference is performed in conjunction with dynamic graph training mechanisms to improve the accuracy of the prediction models and simulate the corresponding model responses when topological relationships or node properties change. The results show that graph deep learning can accurately capture the spatio-temporal characteristics of mixed gas transport processes in a pipeline network. Compared with a normal deep-learning algorithm, the performance improvement of the proposed GNN algorithm can be as high as 62% in R2 and 50% in mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE) metrics, depending on the different scenarios and noises. This model can be applied to various scenarios, such as pipeline extensions and hydrogen blending. Based on all the tests conducted, the generalization ability of the algorithm is better than that of the others, with the accuracies of most cases satisfying the requirements of metering stations.