| 引用本文: | 赵 妍,吴昊鑫,赵宗罗,等.基于深度残差网络和改进时序卷积神经网络的宽频振荡监测[J].电力系统保护与控制,2025,53(24):52-64.[点击复制] |
| ZHAO Yan,WU Haoxin,ZHAO Zongluo,et al.Wideband oscillation monitoring based on deep residual network and improved temporal convolutional neural network[J].Power System Protection and Control,2025,53(24):52-64[点击复制] |
|
| 摘要: |
| 宽频振荡严重威胁电网安全稳定运行。为此，提出一种基于深度残差网络(residual network, ResNet)和改进的时序卷积神经网络(improved temporal convolutional network, ITCN)的宽频振荡监测方法。首先，利用ResNet结构对宽频振荡信号在时间上进行卷积，滑动捕捉时间序列的相邻局部特征，通过残差块的堆叠实现振荡信号多尺度特征的压缩提取。然后，利用ITCN结构通过膨胀因果卷积对压缩特征进行扩展，在保证计算效率的同时逐层引入较大的感受野，进一步提取时间序列中蕴含的中长期依赖特性，两者结合实现了对全局特征的提取。最后，在TCN结构中嵌入注意力机制(Attention)，对信号中重要特征进行加权分配，更好地捕捉全局模式和长期依赖特性。仿真和实测结果验证了ResNet-ITCN模型可以出色地完成宽频振荡参数检测任务并且对振荡类型进行识别，实现了对宽频振荡的监测。 |
| 关键词: 宽频振荡 深度残差网络 改进时序卷积神经网络 注意力机制 滑窗监测 |
| DOI：10.19783/j.cnki.pspc.250075 |
| 投稿时间：2025-01-18修订日期：2025-04-27 |
| 基金项目:国家自然科学基金项目资助(U24B2084)；国网浙江省电力有限公司科技项目资助(5211HZ240001) |
|
| Wideband oscillation monitoring based on deep residual network and improved temporal convolutional neural network |
| ZHAO Yan1,WU Haoxin1,ZHAO Zongluo2,CHEN Yun2,ZHOU Bo2,LI Qiangqiang2 |
| (1. School of Electrical Engineering, Northeast Electric Power University, Jilin 132012, China; 2. State Grid Zhejiang
Electric Power Co., Ltd., Hangzhou Fuyang District Power Supply Company, Hangzhou 310000, China) |
| Abstract: |
| Wideband oscillations pose severe threat to the safe and stable operation of power systems. To address this issue, a wideband oscillation monitoring method based on deep residual network (ResNet) and improved temporal convolutional neural network (ITCN) is proposed. First, the ResNet structure is used to convolve wideband oscillation signals, capturing adjacent local features of the time series through sliding windows. The multi-scale features of the oscillation signals are extracted and compressed by stacking the residual blocks. Then, the ITCN structure applies dilated causal convolutions to expand the compressed features, introducing progressively larger receptive fields while maintaining computational efficiency. This enables further extraction of medium- and long-term dependencies in the time series, and the combination of both networks facilitates comprehensive global feature extraction. Finally, an attention mechanism is embedded into the TCN structure to assign adaptive weights to important signal features, thereby improving the capture of global patterns and long-term dependencies. Simulation and real-world measurements verify that the ResNet-ITCN model can successfully detect wideband oscillation parameters and identify oscillation types, achieving effective wideband oscillation monitoring. |
| Key words: wideband oscillation deep residual network improved temporal convolutional neural network attention mechanism sliding-window monitoring |
