[1]张鹏飞,周恩杰,水薇淋,等.一种基于深度自监督聚类的分布式光伏电网并网功率调节方法[J].机械与电子,2026,44(09):89-95.
　ZHANG Pengfei,ZHOU Enjie,SHUI Weilin,et al.A Deep Self-supervised Clustering-based Power Regulation Method for Distributed Photovoltaic Grid-integration[J].Machinery & Electronics,2026,44(09):89-95.

点击复制

一种基于深度自监督聚类的分布式光伏电网并网功率调节方法()

分享到：

《机械与电子》[ISSN:1001-2257/CN:52-1052/TH]

卷:

44

期数:

2026年09期

页码:

89-95

栏目:

电力控制

出版日期:

2026-09-25

文章信息/Info

Title:

A Deep Self-supervised Clustering-based Power Regulation Method for Distributed Photovoltaic Grid-integration

文章编号:

1001-2257(2026)09-0089-07

作者:

张鹏飞¹; 周恩杰²; 水薇淋²; 潘彦龙¹; 李鹏飞¹

1.国网甘肃省电力公司定西供电公司,甘肃 定西743000;

2.国网甘肃省电力公司定西市安定区供电公司,甘肃 定西 743000

Author(s):

ZHANG Pengfei¹; ZHOU Enjie²; SHUI Weilin²; PAN Yanlong¹; LI Pengfei¹

(1.Dingxi Power Supply Company,State Grid Gansu Electric Power Company,Dingxi 743000,China;

2.Dingxi Anding District Power Supply Company,State Grid Gansu Electric Power Company,Dingxi 743000,China)

关键词:

分布式光伏; 并网功率调节; 自监督学习; 模糊聚类; Transformer

Keywords:

distributed photovoltaic; grid-connected power regulation; deep self-supervised learning; fuzzy clustering; Transformer

分类号:

TM73;TP18

文献标志码:

A

摘要:

为解决分布式光伏并网运行中因多时间尺度波动及缺乏调节标签所导致的控制不稳定问题,提出一种基于深度自监督聚类的并网功率调节方法。该方法利用多时间尺度特征和Transformer网络提取光伏输出的动态表示,并构建基于加权马氏距离的模糊聚类模型,以捕捉运行模式的连续演化特征;同时引入增强视图一致性约束,提升聚类边界的稳定性。将模式嵌入向量与深度表征联合输入策略解码器,通过自监督方式学习调节动作,从而在无显式标签的条件下生成连续、平滑的功率调节指令。仿真结果表明,与传统模糊C均值聚类方法相比,所提方法的类内紧密度降低33.3%,类间分离度提升40.7%,轮廓系数达到0.57,显著优于多种聚类基准方法。在调节性能方面,该方法实现的功率波动标准差为3.18 kW,电压越限率为1.9%,调节代价较基准方法下降13.9%~26.2%。

Abstract:

To address the control instability issues in distributed photovoltaic (PV) systems caused by multi-time scale fluctuations and the absence of supervisory labels,a power regulation method based on deep self-supervised clustering is proposed.The proposed approach extracts the dynamic representation of PV output through multi-time scale features and a Transformer encoder,and constructs a fuzzy clustering model based on the weighted Mahalanobis distance to capture the continuous evolution characteristics of operating patterns.Meanwhile,an enhanced-view consistency constraint is introduced to improve the stability of cluster boundaries.The pattern embedding vectors and deep representations are jointly fed into a strategy decoder,where regulation actions are learned in a self-supervised manner,thereby generating continuous and smooth power-regulation commands without explicit labels.Simulation results demonstrate that,compared with the traditional Fuzzy C Means (FCM) clustering method,the proposed method reduces intra-cluster compactness by 33.3%,increases inter-cluster separability by 40.7%,and achieves a silhouette coefficient of 0.57,significantly outperforming multiple clustering baselines.In terms of regula ation rate of 1.9%,with the regulation cost reduced by 13.9% to 26.2% compared with benchmark approaches.

备注/Memo

备注/Memo:

收稿日期:2025-12-16

基金项目:国网定西供电公司科技项目(B727102500R4)

作者简介:张鹏飞 (1980-),男,甘肃定西人,高级工程师,研究方向电力营销技术、调度自动化技术和分布式光伏控制技术等。

常用功能