ENHANCED MACHINE LEARNING METHODS FOR NONLINEAR HARMONIC SOURCE MODELLING

Power electronic-based devices inject harmonic perturbations in electrical networks and are thus a concern for network operators who must ensure that the harmonic injections and the resulting voltage disturbances are kept within acceptable limits. In this paper, we show that time domain (TD) methods can be useful in preliminary harmonic source modelling to derive the simulated models. Using TD SIMULINK models of our devices, we find the most influential internal and external parameters to be considered when the device is operating in a pre-existing harmonic environment and we generate datasets that can be used for training machine learning (ML) algorithms. We propose here three ML techniques to predict the harmonic currents originating from most widespread power electronic devices such as Light Emitting Diodes (LED) lamps and Electric Vehicle (EV) chargers, and compare them with TD models. The accuracy of all these models regarding their harmonic currents prediction capability is assessed under constantly evolving supply conditions.