电工技术学报  2023, Vol. 38 Issue (17): 4683-4700    DOI: 10.19595/j.cnki.1000-6753.tces.221019
电力系统与综合能源 |
基于残差图卷积深度网络的电网无功储备需求快速计算方法
陈光宇1, 袁文辉1, 徐晓春2, 戴则梅3, 闪鑫3
1.南京工程学院电力工程学院 南京 211167;
2.国网江苏省电力有限公司淮安供电分公司 淮安 223021;
3.南瑞集团公司(国网电力科学研究院) 南京 211167
Fast Calculation Method for Grid Reactive Power Reserve Demand Based on Residual Graph Convolutional Deep Network
Chen Guangyu1, Yuan Wenhui1, Xu Xiaochun2, Dai Zemei3, Shan Xin3
1. School of Electric Power Engineering Nanjing Institute of Technology Nanjing 211167 China;
2. Huai 'an Power Supply Branch of State Grid Jiangsu Electric Power Co. Ltd Huan'an 223021 China;
3. Nanrui Group Co. Ltd State Grid Electric Power Research Institute Nanjing 211167 China
全文: PDF (3741 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 针对电网无功储备需求计算复杂度高、耗时长的问题,提出一种基于残差图卷积深度网络考虑冗余样本特征削减的电网无功储备需求快速计算方法。该文首先,给出一种基于深度学习的电网无功储备需求快速计算框架,采用残差图卷积深度神经网络(GCNII)对电网无功储备需求计算进行建模;其次,为克服传统相似性计算方法在拓扑属性样本度量问题上的局限,提出一种双尺度相似性度量方法,基于矩阵奇异值序列的余弦距离实现对拓扑结构样本的相似性度量;最后,提出一种冗余样本削减策略,基于双尺度相似性度量方法,结合改进谱聚类算法实现对样本集合的分层聚类,并通过样本局部密度分析,实现在维持数据集特征多样性的情况下,对冗余样本进行有效削减,提升模型训练效率。所提算例采用IEEE标准节点系统进行仿真,计算结果表明,该方法能够实现在模型计算精度基本不变的情况下大幅提升模型训练效率。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
陈光宇
袁文辉
徐晓春
戴则梅
闪鑫
关键词 残差图卷积神经网络无功储备需求计算样本削减策略矩阵奇异值序列双尺度相似性    
Abstract:Reactive power reserve plays a crucial role in maintaining voltage stability of power grid. Considering that the uncertainty of new energy output shortens the analysis period of reactive power reserve demand and gradually changes from offline calculation to online evaluation, traditional reactive power reserve demand analysis methods have the problems of high computational complexity and long time consumption. As a result, the calculation of reactive power reserve requirements cannot meet the requirements of online evaluation. To solve these problems, this paper proposes a fast calculation method of grid reactive power reserve demand based on residual graph convolution deep network considering redundant sample reduction. The sample reduction technology and deep learning technology are effectively combined to realize the fast calculation of grid reactive power reserve demand.
Firstly, a fast grid reactive power reserve calculation framework based on deep learning is proposed, and residual graph convolutional neural network (GCNII) is used to model the grid reactive power reserve demand calculation. Secondly, aiming at the limitation of traditional similarity calculation methods in topological attribute sample measurement, a two-scale similarity measurement method was constructed based on feature similarity measure and topological similarity measure. Thirdly, the improved spectral clustering algorithm and densitometric analysis method are combined to deeply mine the redundant data with high similarity and dense distribution in the sample set, and the redundant data are reduced, so as to greatly improve the model training efficiency while ensuring the accuracy of model calculation. Finally, the reduced data sets are used to train and test the deep learning model, and the reactive power reserve requirements of the power grid are rapidly calculated..
The simulation results on IEEE standard system show that the calculation results of the model on the training set and the test set are close to the target value, indicating that the residual graph convolution model has strong generalization ability. Secondly, by comparing the GCNII model with the GCN model at different depths, it can be found that the deep GCNII model has better calculation accuracy, which verifies that the GCNII model can effectively solve the excessive smoothing problem of the GCN model. Finally, the sample reduction strategy is used to effectively process the dataset, and the sample sets before and after the reduction are used to train and calculate the model respectively. It is verified that the sample reduction strategy can greatly improve the model training efficiency while ensuring the model calculation accuracy.
The following conclusions can be drawn from the simulation analysis: (1) Compared with the traditional machine learning model, the GCNII model has higher calculation accuracy, and its mean absolute error (MAE) is 1.777 and 2.779 lower than CNN and SVR models, respectively, indicating that the GCNII model has obvious advantages in reactive power reserve requirement calculation task. (2) The sample reduction strategy can improve the model training efficiency by more than 15%, and the mean absolute error (MAE) of the calculation results is less than 5%, which can greatly improve the model training efficiency while keeping the model calculation accuracy basically unchanged. (3) The simulation results in different node size systems show that the proposed method can effectively improve the model training efficiency of different node sizes, indicating that the proposed method has good universality.
Key wordsResidual graph convolutional neural network    reactive power reserve demand calculation    sample reduction strategy    matrix singular value sequence    two scale similarity   
收稿日期: 2022-06-02     
PACS: TM734  
基金资助:智能电网保护和运行控制国家重点实验室资助(SGNR0000KJJS2302148)
通讯作者: 陈光宇 男,1980年生,博士,副教授,硕士生导师,研究方向为电力系统运行与控制,优化调度,人工智能等。E-mail:cgyhhu@163.com   
作者简介: 袁文辉 男,1997年生,硕士研究生,研究方向为电力系统运行控制,人工智能。E-mail:1272698676@qq.com
引用本文:   
陈光宇, 袁文辉, 徐晓春, 戴则梅, 闪鑫. 基于残差图卷积深度网络的电网无功储备需求快速计算方法[J]. 电工技术学报, 2023, 38(17): 4683-4700. Chen Guangyu, Yuan Wenhui, Xu Xiaochun, Dai Zemei, Shan Xin. Fast Calculation Method for Grid Reactive Power Reserve Demand Based on Residual Graph Convolutional Deep Network. Transactions of China Electrotechnical Society, 2023, 38(17): 4683-4700.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.221019          https://dgjsxb.ces-transaction.com/CN/Y2023/V38/I17/4683