| 摘要: |
| 针对变电站一次接线图建模长期依赖人工、效率低且差错率高的问题，提出一种面向一次接线图的多模型协同智能识别与拓扑建模方法。该方法以改进YOLOv5为图元检测主干，在特征融合阶段引入CARAFE上采样算子，并结合滑窗切片推理提升密集小目标图元的检测能力；采用DB与CRNN构建文本检测与识别链路，实现复杂场景下文本区域的准确提取与内容识别；进一步融合类别先验、空间约束与N-gram语言模型，实现文本—图元匹配与语义纠错；针对母线与连接线结构特征，结合传统图像处理、连通域分析、广度优先搜索及电力拓扑规则，自动生成接线关系并完成一致性校核。在芜湖地区158张一次接线图上的实验结果表明，该方法整体准确率达到94.2%。工程试点显示，新站图模建模平均耗时约2 h/站，效率提升约36倍；存量图模校核平均耗时约0.5 h/站，效率提升约48倍。该方法具有较高准确性与鲁棒性，显著提升建模与校核效率，具备良好的工程应用价值，可为新型电力系统图模数字化建设提供技术支撑。 |
| 关键词: 变电站一次接线图 智能电网 智能识别 多模型协同 |
| DOI： |
| 投稿时间：2026-02-25修订日期：2026-05-12 |
| 基金项目: |
|
| Multi-Model Collaborative Intelligent Recognition and Topology Modeling for Substation Single-Line Diagrams |
|
SONG Mingmin, QIAO Xiangyang, FENG Xipeng
|
| (State Grid Anhui Electric Power Co.,Ltd. Wuhu Power Supply Company) |
| Abstract: |
| To address the long-standing issues of low efficiency and high error rate in substation primary wiring diagram modeling due to reliance on manual methods, this paper proposes a multi-model collaborative intelligent recognition and topology modeling method for primary wiring diagrams. This method uses an improved YOLOv5 as the primitive detection backbone, introduces the CARAFE upsampling operator in the feature fusion stage, and combines sliding window slicing inference to improve the detection capability of dense small target primitives. A text detection and recognition link is constructed using DB and CRNN to achieve accurate extraction and content recognition of text regions in complex scenarios. Furthermore, category priors, spatial constraints, and N-gram language models are integrated to achieve text-primitive matching and semantic error correction. For the structural features of busbars and connecting lines, traditional image processing, connected component analysis, breadth-first search, and power topology rules are combined to automatically generate wiring relationships and complete consistency verification. Experimental results on 158 primary wiring diagrams in the Wuhu area show that the overall accuracy of this method reaches 94.2%. Engineering pilot projects show that the average time for modeling new substation diagrams is approximately 2 hours per substation, representing an efficiency improvement of approximately 36 times; the average time for verifying existing diagram diagrams is approximately 0.5 hours per substation, representing an efficiency improvement of approximately 48 times. This method has high accuracy and robustness, significantly improves modeling and verification efficiency, has good engineering application value, and can provide technical support for the digital construction of new power system diagrams and models. |
| Key words: Substation single-line diagram Smart grid Intelligent recognition Multi-model collaboration |
