基于硬约束和物理信息神经网络结合的天然气管网仿真新方法  

A new gas pipeline network simulation method based on BHC-PINN

作　　者：金凤 郭祎 张岩 马秀云 刘斌[4] 焦健丰 殷雄 徐海龙 宫敬[1] 温凯 JIN Feng;GUO Yi;ZHANG Yan;MA Xiuyun;LIU Bin;JIAO Jianfeng;YIN Xiong;XU Hailong;GONG Jing;WEN Kai(College of Mechanical and Transportation Engineering,China University of Petroleum,Beijing 102249,China;PipeChina Oil and Gas Control Center,Beijing 100000,China;PipeChina Beijing Pipeline Co.,Ltd.,Beijing 100000,China;Oil and Gas Technology Research Institute,PetroChina Huabei Oilfield Company,Renqiu,Hebei 062552,China;Department of Chemical Engineering and Analytical Sciences,University of Manchester,Manchester M13,9PL,UK)

机构地区：[1]中国石油大学(北京)机械与储运工程学院 [2]国家石油天然气管网集团有限公司油气调控中心 [3]国家管网集团北京管道有限公司 [4]中国石油华北油田公司油气工艺研究院 [5]曼彻斯特大学化学工程与分析科学系

出　　处：《天然气工业》2025年第1期164-174,共11页Natural Gas Industry

基　　金：国家自然科学基金面上项目“复杂供气管网大时滞非线性仿真模型构建与智能调控”(编号:52174064);国家自然科学基金项目“面向大规模成品油管网调度的数据解析与优化融合方法”(编号:52202405)。

摘　　要：随着天然气管网互联互通成环成网,需要利用仿真技术模拟各种工况下的管网运行状态,以提升管网的供气能力,进而降低管网运行成本。现有的智能化仿真方法过于依赖数据驱动,忽略了天然气流动机理,导致精度与稳定性较差且结果不可解释。为此,从机理角度分析了天然气流动基本控制方程与管网拓扑结构,将机理信息耦合至深度学习的损失函数指导结构设计与训练,形成了天然气管网物理信息神经网络PINN(Physics-Informed Neural Network)仿真模型;然后描述了输入输出变量之间的耦合关系,设计边界条件嵌入训练硬约束模式,建立了硬约束和物理信息神经网络结合BHC-PINN(Boundary Hard-Constraint Physics-Informed NeuralNetwork)的模型,实现了天然气管网全时空状态的仿真监测。研究结果表明:(1) PINN模型将管道与拓扑结构的机理信息耦合进损失函数以提高模型的可解释性,并借助管网进出口冗余数据提高了准确性,实现了对管网压力流量的准确监测;(2)建立的边界硬约束模式,使神经网络的输出强制满足边界条件,得到的流量和压力仿真误差分别从2.1%和0.32%下降低至1.5%和0.082%,训练的效率和速度分别提高了48.5%和55.9%;(3) BHC-PINN能实现管网任意位置气体流动状态的观测,对于中间阀室压力的仿真最大误差为0.2%。结论认为,该新方法能准确仿真天然气管网内流动状态,增强了数据驱动模型的可解释性,并为天然气管网瞬态仿真提供了数据加机理混合驱动的新思路。With the interconnection and networking of gas pipelines,it is necessary to simulate the operation status of the pipeline network under various working conditions,so as to improve the gas supply capacity of the pipeline network and reduce the operation costs.The existing intelligent simulation methods rely heavily on data drive,but neglect the gas flow mechanism,resulting in poor accuracy and stability.And the results are uninterpretable.In this paper,the basic control equations of gas flow and the topological structure of the pipeline network are analyzed from the perspective of mechanism,the mechanism information is coupled to the loss function of deep learning to guide the structure design and training,and a physics-informed neural network(PINN)simulation model is established.Then,the coupling relationship between the input and output variables is described,the hard constraint mode of embedding boundary condition into training is designed,and a model integrating boundary hard constraint(BHC)and PINN,i.e.BHC-PINN model,is established,so that simulation monitoring of gas pipeline network under the full time-space state is realized.The following results are obtained.First,the PINN model couples the mechanism information of pipelines and topological structures into the loss function to improve the interpretability of the model,and utilizes redundant data of pipeline inlet and outlet to improve accuracy,so as to achieve the accurate monitoring of pipeline pressure and flow rate.Second,the established BHC-PINN model forces the output of the neural network to meet the boundary conditions,resulting in a decrease in flow rate and pressure simulation errors from 2.1%and 0.32%to 1.5%and 0.082%,respectively,and an increase in training efficiency and speed by 48.5%and 55.9%,respectively.Third,the BHC-PINN model can observe the gas flow state at any position in the pipeline network,with the maximum simulation error of the pressure in the intermediate valve chamber being 0.2%.In conclusion,the new BHC-PINN method can accurat

关 键 词：天然气管道 水力瞬态仿真 物理信息神经网络 硬约束模式 

分 类 号：TE832[石油与天然气工程—油气储运工程]