A method of multivariate short-term voltage stability assessment based on heterogeneous graph attention deep network. (March 2022)

• Record Type: Journal Article
• Authors: Zhong, Zhi; Guan, Lin; Su, Yinsheng; Yu, Jingxing; Huang, Jiyu; Guo, Mengxuan
• Is Part Of: International journal of electrical power & energy systems. Volume 136(2022)
• Publication Date: 2022-03
• Subjects: Short-term voltage stability -- Multivariate evaluation index -- Heterogeneous graph attention deep network -- Multi-task learning; Electrical engineering -- Periodicals; Electric power systems -- Periodicals; Électrotechnique -- Périodiques; Réseaux électriques (Énergie) -- Périodiques; Electric power systems; Electrical engineering; Periodicals; 621.3
• DOI: 10.1016/j.ijepes.2021.107648
• Languages: English
• ISSNs: 0142-0615

Abstract

Highlights: The model requires only pre-fault steady state information and the impact data at the fault occurrence period. The method can provide multivariate STVS indices of each bus, and identify the instability patterns. The heterogeneous graph attention network HGAT combine the zoning characteristics with the influence of bus type. The improved multi-tasking learning (MTL) algorithm is proposed. Excellent accuracy and robustness to topological changes. Abstract: Compared with time-domain simulation (TDS), data-driven models show great advantage in time consumption of the power system security analysis. This paper proposes a novel graph neural network (GNN) based model for the short-term voltage stability (STVS) assessment. Based on mainly the steady-state information as the inputs, the model can provide multivariate stability indices for each bus and adapt to minor topological changes. Both the fast voltage collapses (FVC) and fault-induced delayed voltage recovery (FIDVR) events can be identified. A modified multi-task learning (MTL) process improves the multi-index evaluation performance of the model. As the kernel of the model, a heterogeneous graph attention deep network (HGAT) with various types of nodes works to aggregate the local information and generate new features. Then several micro-detectors scan over the network to provide the STVS indices of each bus with only the feature vector of one node as input for the prediction. Comparisons on the test system show that the proposed method can provide fine-grained STVS evaluation and show good robustness in the scenarios of different dynamic load distribution and network topology changes.