电工技术学报  2024, Vol. 39 Issue (16): 4985-4995    DOI: 10.19595/j.cnki.1000-6753.tces.230858
电力系统与综合能源 |
基于时空图卷积网络和自注意机制的频率稳定性预测
杜东来, 韩松, 荣娜
贵州大学电气工程学院 贵阳 550025
Frequency Stability Prediction Method Based on Modified Spatial Temporal Graph Convolutional Networks and Self-Attention
Du Donglai, Han Song, Rong Na
School of Electrical Engineering Guizhou University Guiyang 550025 China
全文: PDF (5566 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 针对传统数据驱动预测方法对电力系统频率稳定性预测的时空特性利用不充分、新拓扑下泛化能力差和可解释性较弱的问题,该文提出了一种基于自注意力机制和时空图卷积网络（STGCN）的频率稳定性预测方法。STGCN预测方法利用一维时间卷积层提取系统时间信息,利用切比雪夫图卷积通过近似拉普拉斯矩阵的多项式函数执行图卷积操作,从而捕获各母线及其邻居的拓扑结构信息;然后,采用基于自注意力机制的可微分图池化层来获得各母线注意力得分以对预测模型的决策过程进行可解释性分析,该分层池化策略允许模型尽可能地保留有价值的节点特征,并根据保留特征和动态拓扑有效分配节点以提高模型的泛化能力与鲁棒性;最后,在修改的新英格兰39节点系统和ACTIVSg500节点系统上的测试验证了所提方法的有效性。与传统方法相比,该文所提STGCN具有更高的预测精度、更好的鲁棒性和泛化能力。同时,该方法可以提供系统内各母线对预测结果的具体影响。
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
杜东来
韩松
荣娜
关键词 ： 频率稳定性预测,  深度学习,  时空图神经网络,  自注意力机制,  可解释性    
Abstract：Power system frequency dynamic characteristics serve for system stability evaluation and reflects the specific change of frequency when there is a power imbalance between generation and load. On the one hand, the distribution of topology is closely connected with the stability of the power grid in the frequency stability prediction (FSP) problem and the topology of the power grid is typically altered by the random events. On the other hand, the frequency stability characteristics are mirrored in the post-fault response trajectory, so that the time-varying features and dynamic topologies may contribute to the nonlinear spatial-temporal dynamics of FSP. However, the traditional data-driven methods fail to effectively incorporate the system spatial-temporal characteristics into the model training, and suffers from insufficient utilization of system information, poor generalization ability in the face of new topology and interpretability. In addition, the machine learning (ML) model employed for prediction resembles a "black box" internally, and lack of interpretability is one of the primary challenges to ML application in the FSP field. To give a highly accurate FSP reference and denote the potential security hazards of the system, the model is required to identify the major factors that influence the FSP and clarify the decision-making process of model learning.
To address these issues, this paper proposes a FSP prediction method that combines the self-attention mechanism (SAM) and the spatial-temporal graph convolutional network (STGCN). Firstly, the proposed STGCN prediction method utilizes a one-dimensional temporal convolutional layer to extract system temporal information. In addition, it employs Chebyshev graph convolution to approximate the Laplacian matrix through polynomial functions, enabling graph convolution operations to capture the topological structure information of each bus and its neighbors. After that, a differentiable self-attention graph pooling (SAGPooling) layer based on SAM is employed to enhance the generalization ability and robustness of the STGCN model. The layer allows the model to reduce the dimensionality of the feature vectors in order to decrease the number of parameters and avoid overfitting. The hierarchical pooling strategy enables the model to preserve valuable node features as much as possible and effectively allocate nodes based on the preserved features and changing topology to enhance the generalization ability and robustness of the STGCN. Meanwhile, the attention scores of each node can be uniformly extracted. Finally, through the SAM, the attention scores of nodes are obtained according to the active power to perform the interpretability analysis of the STGCN model. In summary, this model converts input data into high-level representations of graphics through graph convolution, time convolution, and SAGPooling to integrate the complete spatiotemporal dynamics of FSP. Therefore, the accuracy, generalization ability, and robustness of the proposed STGCN have been improved, and the interpretable analysis of the model decision-making process can be carried out. The testing results on the modified New England 39-bus system and the modified ACTIVSg500 system, which incorporate renewable energy sources, validate the effectiveness of the proposed STGCN. Among all the tested methods, the STGCN has higher prediction accuracy, better robustness, and generalization capability. In addition, the STGCN can provide critical influence factors of different buses on the prediction results in this work.
Key words： Frequency stability prediction    deep learning    spatial-temporal graph neural networks    self-attention mechanism    interpretability analysis   
收稿日期: 2023-06-06     
PACS: TM712  
基金资助:贵州省优秀青年科技人才项目（[2021]5645）、贵州省科技支撑计划项目（[2023]290, [2023]329）和贵州省科学技术基金项目（[2021]277）资助
通讯作者: 韩 松 男,1978年生,教授,研究方向为交直流电力系统分析、新型电力电子装备以及配网规划。E-mail：shan@gzu.edu.cn   
作者简介: 杜东来 男,1998年生,硕士研究生,研究方向为人工智能技术在电力系统频率稳定性中的应用。E-mail：1640371942@qq.com
引用本文:   
杜东来, 韩松, 荣娜. 基于时空图卷积网络和自注意机制的频率稳定性预测[J]. 电工技术学报, 2024, 39(16): 4985-4995. Du Donglai, Han Song, Rong Na. Frequency Stability Prediction Method Based on Modified Spatial Temporal Graph Convolutional Networks and Self-Attention. Transactions of China Electrotechnical Society, 2024, 39(16): 4985-4995.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.230858          https://dgjsxb.ces-transaction.com/CN/Y2024/V39/I16/4985