Dropout Kolmogorov–Arnold Networks: A Novel Data-Driven Impedance Modeling Approach for Voltage-Source Converters

Abstract

The extensive integration of voltage-source converters (VSCs) as interfaces for renewable energy sources in power systems increases stability concerns and demands accurate VSC impedance models to characterize grid-converter interactions at various operating points. However, analytical impedance models require detailed knowledge of the VSC parameters, which are frequently inaccessible due to manufacturer confidentiality. Further, existing neural network data-driven VSC impedance identification methods adopt conventional multi-layer perceptrons, yielding complex models and demanding abundant high-quality data. This paper presents a data-driven VSC impedance identification method using Dropout Kolmogorov-Arnold Networks (DropKANs) to address these challenges effectively. The hyperparameters of the proposed DropKAN model are optimized using Optuna, outperforming the Scikit-learn, Hyperopt, and GPyOpt optimizers, and the training is optimized using the Adam optimizer and compared with Nadam and RMSprop. Comprehensive and comparative evaluation tests showed 1) the superiority of the proposed DropKAN model over the feedforward neural network, long short-term memory, and KAN models in terms of accuracy, training and prediction times, and neural network structure simplicity, even with a 50% reduction in the training data size, and 2) the versatility and robustness of the proposed DropKAN model when applied to a different VSC system.