基于级联神经网络的变压器复杂异常录波波形智能识别及恢复方法

An Intelligent Method for Recognition and Restoration of Transformer Complex Abnormal Waveforms Based on Cascaded Neural Networks

中文关键词:异常波形分析  变压器  电流互感器  励磁涌流  卷积神经网络

keywords:abnormal waveform analysis  transformers  current transformers  inrush current  convolutional neural networks

中文摘要:针对电力变压器和电流互感器（current transformer， CT）在铁心饱和时产生的复杂异常波形分析难度大的问题，提出了一种基于级联神经网络的变压器复杂异常录波波形智能识别及恢复方法，能够高效判别异常类型并准确提取其基波分量。首先，建立变压器励磁涌流、CT饱和的机理模型，分析其异常波形的主要影响因素；其次，构建级联卷积神经网络（convolutional neural network，CNN）框架，包含异常波形分类的一级神经网络，以及励磁涌流、CT饱和电流的非饱和区识别的二级神经网络，通过神经网络的级联实现波形分析任务的分解；再次，基于波形非饱和区识别结果，采用最小二乘法对励磁涌流和CT饱和电流波形进行恢复，得到其基波分量；最后，通过仿真波形对所提方法进行性能验证。结果表明，所提方法相比传统方法显著提升了异常录波分析的效率和准确性。

Abstract:This paper addresses the challenge of analyzing complex abnormal waveforms generated by power transformers and current transformers （CT） during core saturation. It proposes an intelligent recognition and restoration method for transformer complex abnormal recorded waveforms based on cascaded neural networks. The proposed method enables rapid and efficient identification of abnormal types from abnormal recordings and accurately extracts their fundamental components. First，a mechanism model for transformer inrush currents and CT saturation is established to analyze key influencing factors of these abnormal waveforms. Subsequently，a cascaded convolutional neural network （CNN） framework is constructed，featuring a primary network for abnormal waveform classification and a secondary network dedicated to identifying unsaturated regions within inrush and CT saturation currents. This cascading architecture effectively decomposes the waveform analysis task. Next，based on the identified unsaturated segments of the waveforms，the inrush and CT-saturation currents are restored using the least-squares method，yielding their fundamental components. Finally，the method’s performance is rigorously verified using simulated waveforms. The results demonstrate that the proposed method significantly enhances the efficiency and accuracy of abnormal recording analysis compared to conventional approaches.