Research On Calorific Value Assignment Method Based On Machine Learning For Natural Gas Pipeline Networks With Comingled Gas Flows

The rapid development of China’s natural gas industry has accelerated the construction of the energy determination system for natural gas.Accurate and reliable calorific value analysis and assignment technology is an important basis for the energy determination of natural gas pipeline networks.Because the diversified gas supply and complex operating environment,the transmission with comingled gas flows lead to sharp fluctuation of the calorific value in the networks.It can hardly meet the needs for realtime and accurate calorific value assignment at different interfaces of the networks by the existing determination process and traditional analysis techniques.Therefore,this research aims to explore that how to use the machine learning to support the real-time calorific value analysis in the networks,in which we study the calorific value assignment method from the change of the topology and the operating conditions of the pipeline networks,and improve the transfer ability of the method.Firstly,an accurate natural gas calorific value prediction method is proposed to solve the issue that gas quality changes caused by the comingling of multiple gas flows.This method introduces a calculation model of physical parameters,such as the calorific value and compressibility factor,that based on the semi-empirical equation by integrating the gas flow state,transport characteristic and thermodynamic property.Then,a calorific value prediction model is developed using the machine learning algorithm to mine the mapping pattern between the flow parameters and calorific value of the comingled gas flows.The accuracy and adaptability of the machine learning model is compared with the semi-empirical calculation model.Case study shows that the accuracy of this method is99.5%,which is robust with the different data quantities and quality.It can be applied to the small branched networks and could provide a reference for real-time gas quality analysis.Secondly,considering the influence of various factors on the gas state,such as the feature of gas transmission process,the variation of gas supply and the fluctuation of demand,a deep learning-based prediction model is developed to address the complex calorific value fluctuation problem in the large-scale and multi-sources natural gas pipeline networks.The model combines the dynamic characteristics of the system with large time delay and nonlinearity,which can be used to analyze the dynamic change of calorific value caused by unsteady process in the networks accurately and efficiently.The numerical experiments show that the deep learning model can extract the influence of the unsteady and large-time-delay hydraulic characteristics of the networks on the gas calorific value profile.This method can predict the dynamic change of calorific value in the networks based on real-time operational data,such as pressure,flow rate and gas quality parameters,with a prediction accuracy over 99%.The computational efficiency of this model is improved by 99% compared to that of the physical simulation model.Furthermore,this method has a relatively high accuracy over 91% with noisy datas and missing crucial parameters,such as pressure and flow rate of some grid nodes,which can provide a new pathway for the dynamic assignment of calorific value in the complex gas pipeline networks and field metering management.Finally,considering the topological structure and the dynamic characteristic of the system,a spatio-temporal prediction model based on the graph deep learning theory is developed,in order to overcome the degradation of prediction performance of the deep learning model under different data conditions and extended analysis scenarios in the natural gas pipeline networks.This model can capture the spatio-temporal evolution characteristic of the gas state accurately in the networks,which can be transferred to new analysis scenarios by the inductive reasoning mechanism of the GNN.The results show that this method can improve the accuracy of prediction model effectively under different prediction tasks and noisy datas,whose performance can be improved by up to 62%compared to the former model.Besides,it is scalable and can be transferred to new analysis scenarios,such as topology change and hydrogen injection of the networks.