融合配电台区多元特征的轻量化CenterNet设备识别方法
An Equipment Identification Method Based on Lightweight CenterNet from Multivariate Features Fusion for Distribution Area
王文彬 1范瑞祥 1邓志祥 1万军彪 1潘建兵1
作者信息
- 1. 国网江西省电力有限公司电力科学研究院,南昌 330096
- 折叠
摘要
对配电台区设备进行识别,不仅为电网升级提供详细数据,还可以有效地提高配电网低压运行水平.针对配电台区负荷设备提出一种识别方法,融合负荷设备的拓扑结构特征、电力时序数据稳态特征、设备精度寿命特征多元特征值进行设计,在边缘设备处部署CenterNet算法,实现无监督学习的配电台区设备识别.考虑到嵌入式边缘设备计算能力不足,对算法采用多种轻量化处理,包括减小输入数据的尺寸、对网络及推理结构进行优化、剪枝等操作,使得数据检测速度从1720 ms提高到322 ms,达到了实时处理的目标,为后续配电台区设计及优化提供了一种有效手段.
Abstract
Identifying the equipment in the distribution station area is necessary.It can not only provide detailed grid connection data for grid upgrading and new energy,but also effectively improve the low-voltage operation level of the distribution network.In this paper,a method of equipment identification in the distribution station area is proposed,which combines the topological structure characteristics of the equipment in the distribution station area,the steady-state characteristics of the power time series data and the lifetime accuracy characteristics of the equipment in the distribution station area.The CenterNet algorithm is deployed at the edge equipment to realize unsupervised learning for equipment identification in the distribution area.Considering the lack of computing power of embedded devices,the CenterNet algorithm adopts a variety of lightweight processing,including reducing the size of input data,optimizing the network structure and reasoning structure,pruning and other operations,so that the data detection speed is increased from 1720 ms to 322 ms,which can achieve the goal of real-time processing on the premise of meeting the accuracy requirements and provides an effective means for the subsequent design and optimization of the distribution station area.
关键词
CenterNet/配电台区/非监督学习/轻量化Key words
CenterNet/distribution area/unsupervised learning/lightweight引用本文复制引用
出版年
2024
NETL
NSTL
万方数据