数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (2): 161-174.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.02.016

doi: 10.11871/jfdc.issn.2096-742X.2025.02.016

• 技术与应用 • 上一篇    下一篇

结合ICEEMDAN与融合分类模型的宽频振荡辨识方法

李朋波*(),朱晓峰,沈明慷,沙浩源,何茂慧,邓凯,许瑨   

  1. 国网江苏省电力有限公司,超高压分公司,江苏 南京 211102
  • 收稿日期:2024-09-30 出版日期:2025-04-20 发布日期:2025-04-23
  • 通讯作者: 李朋波
  • 作者简介:李朋波,硕士研究生,高级工程师,研究方向为电力系统自动化。
    本文负责文章模型的构建,主要内容的撰写。
    LI Pengbo is a master’s student and senior engineer. His main research direction is power system automation.
    In this paper, he is responsible for the model construction and the writing of the main content.
    E-mail: peng_bo_li@163.com
  • 基金资助:
    国网江苏省电力有限公司科技项目(J2023144)

A New Method for Identification of Broadband Oscillations Based on ICEEMDAN and Fusion Classification Model

LI Pengbo*(),ZHU Xiaofeng,SHEN Mingkang,SHA Haoyuan,HE Maohui,DENG Kai,XU Jin   

  1. EHV Branch Company, State Grid Jiangsu Electric Power Company, Nanjing, Jiangsu 211102, China
  • Received:2024-09-30 Online:2025-04-20 Published:2025-04-23
  • Contact: LI Pengbo

摘要:

【目的】为提高电力系统宽频振荡辨识的准确率,本文提出一种融合ICEEMDAN-KPCA-KAN的宽频振荡辨识方法。【方法】首先,对于随机性、波动性强的原始序列,采用改进的自适应噪声完全集合经验模态分解(ICEEMDAN),将原始宽频振荡信号分解成多个本征模态分量(IMFs)与残差分量,以减小模态混叠现象、凸显各分量时频域特性。其次,针对分量数目多、总特征维数高存在冗余的问题,采用皮尔逊相关系数法,从这些IMFs中筛选出与原始信号相关性最强的5个IMF。进一步地,针对每个IMF,提取其在时域和频域的特征参数,根据相关系数的大小,将特征最终确定为其时频域特征或其本身。针对特征为其本身的IMFs,采用CNN与LSTM融合模型作为分类器;针对特征为其时频域的IMFs,使用核主成分分析(KPCA)技术,对这些高维特征向量进行非线性降维,形成适用于电力系统宽频振荡辨识的特征集,并将降维后的低维向量送入Kolmogorov-Arnold网络(KAN)进行分类辨识。最终,使用LightGBM模型整合各分类器输出,生成融合分类概率,得到最终的分类结果。【结论】通过仿真算例测试表明,该融合方法能够快速、准确地辨识的宽频振荡信号,辨识率达到99.73%。

关键词: 宽频振荡辨识, ICEEMDAN, KPCA, Kolmogorov-Arnold网络

Abstract:

[Objective] To enhance the accuracy of wideband oscillation identification in power systems, this paper proposes a method that integrates ICEEMDAN-KPCA-KAN for wideband oscillation identification. [Methods] Firstly, for the original sequence with strong randomness and volatility, we use ICEEMDAN to decompose the original wideband oscillation signal into multiple intrinsic mode functions (IMFs) and a residual component, thereby reducing modal aliasing and highlighting the time-frequency characteristics of each component. Secondly, For each IMF, we extract time and frequency domain features and determine whether to use the IMF itself or its time-frequency features based on correlation. IMFs used directly are classified using a combined CNN-LSTM model. For those that need time-frequency features, we use KPCA to reduce the dimensions of high-dimensional data and create a feature set. This reduced data is fed into the Kolmogorov-Arnold Network (KAN) for classification. Finally, a LightGBM model combines the outputs from both classifiers to produce the final classification result. [Conclusions] Simulation examples demonstrate that this integrated method can quickly and accurately identify wideband oscillation signals with an identification rate of 99.73%.

Key words: wide-band oscillation identification, ICEEMDAN, KPCA, Kolmogorov-Arnold