A systematic hybrid method for real-time prediction of system conditions in natural gas pipeline networks

The current framework of management of natural gas pipeline systems, based on off-line simulation, is facing challenges because of the increasing complexity, uncertainty and a number of time-dependent factors. To be effective, it requires comprehensive knowledge of system characteristics, accurate initial and boundary conditions. In an attempt to circumvent these problems, in this work we propose to use the deep learning method in the natural gas transmission system operation and management context. A data-driven prediction method is developed from real-time data of operation pressure and gas consumption. Specifically, the deep learning method is combined with the data window method and structural controllability theory to predict the conditions of gas pipeline network components. The data window method is applied to reconstruct the data structure and build a “memory” for the deep learning method. Structural controllability theory is applied to extract critical parameters, for reducing the problem size. The developed method allows accurate and efficient predictions, especially in abnormal conditions. For demonstration, the method is applied to a complex gas pipeline network. The results show that the developed method can provide accurate real-time predictions useful for reducing potential losses in operation, and perform efficient and effective management of the gas pipeline system. In the case study, the average prediction accuracy is higher than 0.99.