基于物理信息神经网络的热网动态状态估计方法  

Dynamic state estimation method of heating network based on physics-informed neural networks

作　　者：张佳琛 郭庆来[1] 王志伟 孙勇 李宝聚 尹冠雄 孙宏斌[1] ZHANG Jiachen;GUO Qinglai;WANG Zhiwei;SUN Yong;LI Baoju;YIN Guanxiong;SUN Hongbin(Department of Electrical Engineering,Tsinghua University,Beijing 100084,China;State Grid Jilin Electric Power Company,Changchun 130012,China)

机构地区：[1]清华大学电机工程与应用电子技术系,北京100084 [2]国网吉林省电力有限公司,吉林长春130012

出　　处：《电力自动化设备》2023年第10期69-78,共10页Electric Power Automation Equipment

基　　金：国家重点研发计划项目(2022YFB2404000)。

摘　　要：在城市综合能源系统中,热网状态估计针对慢动态系统,存在计算精度低、参数不准确、量测不完备的特点。基于物理信息神经网络(PINNs),将含偏微分方程约束的热网动态状态估计问题转化为自动满足偏微分方程约束的神经网络训练问题,并基于损失函数对参数的梯度下降完成热网参数的在线辨识;再将其应用于滚动时间窗中进行在线训练,实现了状态量的动态追踪;进一步基于PINNs对未来时间窗的预测能力提出了一种新的坏数据辨识方法;最后在5节点和27节点热网算例中验证了所提方法的有效性。In the urban integrated energy system,in view of the slow dynamic system,the state estimation of the heating network is characterized by low calculation precision,inaccurate parameters,and incomplete measurements.The problem about dynamic state estimation of heating network with partial differential equa⁃tion constraints is transformed into the training problem of neural networks that automatically satisfy the constraints based on physics-informed neural networks(PINNs).Besides,the gradient descent of the loss function to parameters can be used to achieve the online identification of heating network’s parameters.Then,it is applied to online training in a rolling time window to realize dynamic tracking of state variables.Furthermore,a new method for identifying bad data is proposed based on the predictive ability of PINNs in the future time period.Finally,the effectiveness of the proposed method is verified by the 5-node and 27-node heating network examples.

关 键 词：状态估计 热动态 物理信息神经网络 模型-数据驱动 

分 类 号：TM73[电气工程—电力系统及自动化]