| 引用本文: | 余 明,姚 伟,赵一帆,等.基于短时傅里叶变换和两阶段深度迁移学习的多频段振荡源定位[J].电力系统保护与控制,2025,53(03):81-94.[点击复制] |
| YU Ming,YAO Wei,ZHAO Yifan,et al.Multi-frequency band oscillation source location based on STFT andtwo-stage deep transfer learning[J].Power System Protection and Control,2025,53(03):81-94[点击复制] |
|
| 摘要: |
| 随着以风光为代表的新能源发电大规模并入电力系统以及大型水轮机组的新型调速器的投入使用，新型电力系统的振荡从传统的低频振荡扩展到多频段振荡，准确定位多频段振荡源是抑制不利影响扩大的关键手段。基于此，提出了一种基于短时傅里叶变换和两阶段深度迁移学习的定位方法。首先，该方法将系统中发电机的有功量测信号通过短时傅里叶变换处理转换得到时频表征矩阵，并通过线性映射将其转化为特征图像，从而将定位问题转化为图像分类问题。然后，将所得到的图像输入到基于ResNet50的两阶段分类器。第一阶段用于确定振荡的类型，第二阶段则用于确定该类型振荡源的具体位置。采用融入图像知识学习的迁移学习进一步提高训练效率和定位准确率。含风电的新英格兰系统和湖北电网算例仿真结果均表明：相较于支持相量机、决策树和单阶段迁移学习方法，所提方法在面对噪声时兼顾较高的准确性和鲁棒性。 |
| 关键词: 多频段振荡 短时傅里叶变换 特征图像 深度迁移学习 振荡源定位 |
| DOI：10.19783/j.cnki.pspc.240251 |
| 投稿时间：2024-03-05修订日期：2024-05-22 |
| 基金项目:国家电网公司总部科技项目资助“新一代人工智能技术在未来电网安全分析与决策中的应用”(5100- 202099522A-0-0-00) |
|
| Multi-frequency band oscillation source location based on STFT andtwo-stage deep transfer learning |
| YU Ming1,YAO Wei1,ZHAO Yifan1,SHI Zhongtuo1,LIU Haiguang2,CHEN Rusi2,LI Dahu3,WEN Jinyu1 |
| (1. School of Electrical and Electronic Engineering, Huazhong University of Science and Technology,
Wuhan 430074, China; 2. State Grid Hubei Electric Power Research Institute, Wuhan 430077, China;
3. State Grid Hubei Electric Power Co., Ltd., Wuhan 430077, China) |
| Abstract: |
| With the large-scale integration of new energy, represented by wind and solar power, into the system and the use of new governors of high-power turbine units, the oscillations in the new power system have expanded from the traditional low frequency oscillation to multi-frequency band oscillation. Accurately locating the oscillation source is a key means to suppress the expansion of adverse effects. Thus a novel location method based on short-time Fourier transform (STFT) and two-stage deep transfer learning is proposed. In this method, the active power measurement signals of all generators are converted into time-frequency representation matrices by STFT processing, and the matrices are transformed into feature images by linear mapping, so that the location problem is transformed into an image classification problem. The feature images are then fed into a ResNet50-based two-stage classifier. The first stage is used to determine the type of oscillation, while the second stage is used to locate the source. Transfer learning integrated with image knowledge learning is adopted to further improve the training efficiency and localization accuracy. Simulation results for the New England system with wind power and the Hubei power grid show that, compared to the support vector machine, decision tree and single-step transfer learning method, the proposed method has higher accuracy and robustness in the presence of noise. |
| Key words: multi-frequency band oscillation short-time Fourier transform (STFT) feature image deep transfer learning oscillation source location |
