摘要
高频传输线电压的高效建模与预测对电力系统和通信领域的稳定运行至关重要。然而,现有物理信息神经网络(PINNs)受时间离散化采样策略的限制,可能导致信息损失,影响预测精度。为此,提出一种循环回溯PINNs来捕捉时间依赖性,以增强小样本数据的信息表达能力,从而提高预测精度。首先,构建包含两个核心子网络的循环回溯神经网络架构:网络1负责初步预测电压,网络2基于网络1的预测结果,通过循环回溯机制来补偿因时间离散化导致的信息损失;然后,融合上述两个网络,形成数据-物理信息融合损失函数,在学习数据特征的同时遵循物理规律;最后,通过数值仿真实验验证了所提方法在小样本条件下保持了较高的预测精度,并实现了波速辨识。
Efficient modeling and prediction of high-frequency transmission line voltage are crucial for the stable operation of power systems and communication fields.However,existing Physics-Informed Neural Networks(PINNs)are limited by time discretization sampling strategies,which may result in information loss and affect prediction accuracy.To this end,a recurrent backtracking PINNs is proposed to capture temporal dependencies,enhance the information expression ability of small sample data,and thus improve prediction accuracy.Firstly,construct a recurrent backtracking neural network architecture consisting of two core subnetworks:Network 1 is responsible for initial voltage prediction,while Network 2 compensates for information loss caused by time discretization based on the prediction results of Network 1 through a recurrent backtracking mechanism.Then,the above two networks are fused to form a data-physics-hybrid loss function,which learns data features while following physical laws.Finally,numerical simulation experiments are conducted to verify that the proposed method maintained high prediction accuracy under small sample conditions and achieved wave velocity identification.
作者
李通博
黄浩
迟俊鑫
赵洋
黄海鸣
LI Tongbo;HUANG Hao;CHI Junxin;ZHAO Yang;HUANG Haiming(Liaoning Electric Power Survey&Design Institute Co.,Ltd.,China Energy Engineering Group,Shenyang,Liaoning 110176,China;Liaoyang Petroleum Steel Pipe Manufacturing Co.,Ltd.,Liaoyang,Liaoning 111003,China)
出处
《自动化应用》
2026年第2期158-163,166,共7页
Automation Application
