Dynamic simulation of natural gas pipeline network based on interpretable machine learning model

Abstract

Natural gas pipeline network modeling and simulation is the basis of dispatch and design. Modeling methods based on the mechanistic model have for a long time been facing the problem of multi-parameters and multi-flow patterns that are difficult to determine. Additionally, the method of purely machine learning has the problems of poor interpretability and difficulty in optimizing the model. A novel dynamic simulation method based on an interpretable shortcut Elman network (Shortcut-ENN) model for the pipeline network is proposed. The Shortcut-ENN model is derived from the state space equations. Based on the Shortcut-ENN model, the connection relationship and mechanism characteristics of the pipeline are retained, and an interpretable machine learning pipeline network model is constructed to make up for the lack of mechanism modeling. The model fully adopts the mechanism knowledge and is very suitable for optimization, which greatly improves robustness of the model. Validated and compared with long short-term memory model, the results show that MSE, MAE, R2, and EV of the proposed Shortcut-ENN-based model considering embedded pipeline mechanism and compressor constraints are improved approximately 84.4%, 60.1%, 0.75%, and 53.3%, respectively, and the R2 is about larger than 0.99, and the EV is about less than 0.02.