| 引用本文: | 郑博妍,袁 至,李 骥.基于神经盲反卷积时频域去噪的电能质量扰动联合识别方法[J].电力系统保护与控制,2025,53(21):50-61.[点击复制] |
| ZHENG Boyan,YUAN Zhi,LI Ji.Joint identification method for power quality disturbances based on neural blind deconvolution time-frequency domain denoising[J].Power System Protection and Control,2025,53(21):50-61[点击复制] |
|
| 摘要: |
| 现有的电能质量扰动(power quality disturbances, PQDs)去噪算法存在易丢弃真实信号、去噪效果差、无法识别1/f噪声和Laplace噪声等缺点。为提高噪声下PQDs识别精度和效率，提出一种基于神经盲反卷积(neural blind deconvolution, NBD)时频域去噪的PQDs联合识别方法。首先，构建由NBD与Transformer组成的联合识别模型， NBD整合了时域二次卷积滤波器和频域线性滤波器以实现去噪功能。Transformer负责从去噪后数据中提取特征并进行分类任务。其次，为保证训练效果最优，提出了基于贝叶斯不确定性的动态加权策略，构建由峰度、包络谱目标函数和交叉熵损失组成的联合损失函数对所提模型进行优化。最后，基于IEEE Std 1159-2019标准生成25类PQDs并进行仿真实验。仿真结果表明，该方法实现了不同噪声类型下PQDs的准确识别，相较于其他方法具备更优的F1分数、Params、FLOPs等，提高了去噪性能、识别精度和效率。 |
| 关键词: 神经盲反卷积 时域二次神经滤波器 频域线性神经滤波器 时频域去噪 联合损失函数 电能质量扰动 |
| DOI：10.19783/j.cnki.pspc.250041 |
| 投稿时间：2025-01-13修订日期：2025-04-16 |
| 基金项目:新疆维吾尔自治区重大科技专项资助(2022A01001-4) |
|
| Joint identification method for power quality disturbances based on neural blind deconvolution time-frequency domain denoising |
| ZHENG Boyan1,YUAN Zhi1,LI Ji2 |
| (1. Engineering Research Center for Renewable Energy Power Generation and Grid-connected Control, Ministry of
Education, Xinjiang University, Urumqi 830017, China; 2. Electric Power Research Institute,
State Grid Xinjiang Electric Power Co., Ltd., Urumqi 830011, China) |
| Abstract: |
| Existing denoising algorithms for power quality disturbances (PQDs) have shortcomings such as loss of true signal components, poor denoising performance, and inability to identify 1/f noise and Laplace noise. In order to improve the accuracy and efficiency of PQD identification under noisy conditions, a joint identification method based on neural blind deconvolution (NBD) time-frequency domain denoising is proposed. First, a joint identification model combining NBD and Transformer is constructed. The NBD integrates a time-domain quadratic convolution filter and a frequency domain linear filter to realize denoising, while the Transformer is responsible for extracting features and performing classification on the denoised data. Second, to ensure optimal training effect, a dynamic weighting strategy based on Bayesian uncertainty is proposed, and a joint loss function composed of kurtosis, envelope spectrum objective function, and cross-entropy loss is introduced to optimize the proposed model. Finally, based on the IEEE Std 1159-2019 standard, 25 classes of PQDs are generated and simulated. The simulation results show that the proposed method achieves accurate identification of PQDs under different noise types, and outperforms other methods in terms of F1 score, Params, and FLOPs, thereby improving denoising performance, identification accuracy and computational efficiency. |
| Key words: neural blind deconvolution time domain quadratic neural filter frequency domain linear neural filter time-frequency domain denoising joint loss function power quality disturbance |
