Graph attention-based U-net conditional generative adversarial networks for the identification of synchronous generation unit parameters

Abstract

Listen

Translate article

After generation units are connected to the power grid, electrical parameters such as generation unit output voltage, current and power need to be monitored for the safe and reliable function of grids. Obtaining accurate parameters of synchronous generation units (SGUs) is normally the basic requirement for grid stability analysis and control. To improve the accuracy of identifying SGU parameters, this study proposes graph attention-based U-Net conditional generative adversarial networks (GAUCGANs) for the identification of SGU parameters. The mapping between real and generated samples is modeled by applying conditional generative adversarial networks (CGANs). The CGANs of the GAUCGANs proposed in this study consist of U-Nets-based generators, fully convolutional networks-based discriminators, and graph attention networks (GATs). The U-Nets of the GAUCGANs enhance the ability of generators to generate generative samples that are infinitely close to real samples. The U-Nets-based generators of the GAUCGANs are responsible for generating the SGU parameters; the fully convolutional networks-based discriminators of the GAUCGANs determine whether data is from the generator or the real data. In addition, the classification loss and sample labels are derived from the GATs. The addition of GATs will further enhance the ability of the generators to remove adversarial disturbances. In this study, the proposed GAUCGAN for parameter identification is compared with the traditional parameter identification algorithms. The comparison algorithms are applied to the Simulink standard SGU, hydroelectric unit No. 6 of the waterfall gorge hydropower plant, and the generation unit at the 18th bus of the IEEE 36-bus power system. The numerical simulation results from the GAUCGANs verify that the GAUCGAN can more accurately identify SGU parameters than the traditional parameter identification algorithms. The GAUCGANs improve the accuracy by at least 61.60% on average over the other methods by comparing the RMSE, MAE, MAPE, and SMAPE under three cases comprehensively.