摘要
油中溶解气体分析(DGA)是变压器状态监测和故障预警的核心方法。然而,传统基于数据驱动的DGA预测方法在准确性和泛化能力方面存在一定局限性。为了解决上述问题,提出一种基于物理信息神经网络(PINN)的变压器油中溶解气体预测方法。首先,构建典型缺陷模型的油纸绝缘实验平台,研究局部放电下DGA产气规律,并建立溶解气体与放电能量的关联模型。然后,构建物理信息约束方程组,利用该方程组嵌入时序预测模型并进行训练,以提高预测精度和可解释性。实验结果表明,相较于传统智能模型,PINN模型能够更准确地预测油中溶解气体浓度变化,并在数据稀缺的情况下仍能保持较高的稳定性和泛化能力。研究为提升变压器健康状态监测的智能化水平提供了一种新思路,有助于实现更精准的状态评估和早期故障预警。
Dissolved gas analysis(DGA)is the core method for transformer condition monitoring and early fault warning.However,traditional data-driven DGA prediction methods have some limitations in accuracy and generalization.To address these limitations,a prediction method for dissolved gases in transformer oil based on a physics-informed neural network(PINN)was developed.An oil-paper insulation experimental platform with typical defect models was constructed to investigate the gas generation behavior of partial discharges in DGA,and a correlation model between dissolved gases and discharge energy was established.Physical information constraint equations were then formulated and embedded into a time-series prediction model,which was trained to enhance prediction accuracy and interpretability.Experimental results indicate that,compared with traditional intelligent models,the proposed PINN model can predict the variation of dissolved gas concentrations in oil more accurately and can maintain high stability and generalization capability under data-scarce conditions.This study provides a new approach to improve the intelligence level of transformer health condition monitoring and supports more accurate state assessment and early fault warning.
作者
董明
陈骥
常昊鑫
胡一卓
张崇兴
张海滨
董璇
DONG Ming;CHEN Ji;CHANG Haoxin;HU Yizhuo;ZHANG Chongxing;ZHANG Haibin;DONG Xuan(State Key Laboratory of Electrical Insulation and Power Equipment,Xi’an Jiaotong University,Xi’an 710049,China;NARI Technology Co.,Ltd.,Nanjing 211106,China)
出处
《电机与控制学报》
北大核心
2025年第12期1-12,共12页
Electric Machines and Control
基金
国家重点研发计划(2023YFB2406900)。
关键词
电力变压器
油中溶解气体预测
物理信息神经网络
产气规律
人工智能
power transformer
prediction of dissolved gas in oil
physics-informed neural networks
gas production law
artificial intelligence
