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

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.