Machine Learning-Enhanced Converter Models for Harmonic Analysis in Distribution Grids

ABSTRACT This paper proposes a methodology that integrates machine learning techniques into harmonic load flow analysis for electrical distribution networks. Nonlinear loads are represented as black-box models, where the harmonic components of the load currents are estimated based on the frequency spectrum of the supply voltage and the equivalent impedance. This eliminates the need for prior knowledge of internal architecture or parameters of the loads, including power converters and control algorithms. The proposed workflow encompasses dataset construction, data preprocessing, and the training and evaluation of three machine learning models using appropriate performance metrics. Unlike conventional methods, the proposed approach explicitly captures the interactions between converter or nonlinear load currents and grid operating conditions. The methodology is validated using a typical CIGRE low-voltage benchmark network, where the apparent power of the machine learning based model is progressively increased. During this process, numerical convergence challenges arise and are effectively addressed through an alternative mathematical framework describing the coupling between the harmonic load flow and machine learning models. This reformulation not only guarantees convergence but also significantly improves computational efficiency. A comparison with time-domain simulations confirms that the frequency-domain results converge toward the physical solution within an acceptable error margin. This demonstrates the robustness and accuracy of the proposed approach, highlighting its potential to overcome key limitations of existing harmonic analysis models.