Machine Learning Metamodeling of Harmonic Sources in LV Distribution Networks

Power electronic-based non-linear devices inject harmonic current into the distribution grids they are connected to. Their injection depends on the preexisting distortions of their supply voltage and on the grid upstream impedances. Electromagnetic Transient (EMT) techniques capture quite accurately the non-linear relationships between the network and the device harmonic injections thanks to detailed time models. Their downsides are that detailed power electronic models are seldom available to grid operators, and when available, the simulations are time- and resource-consuming as new EMT simulations are required whenever the network conditions evolve. In recent years, machine learning metamodels (MLM) have been able to accurately predict the behavior of non-linear systems. We apply this approach to model harmonic sources in the frequency-domain and present in this paper how MLM accurately predict the harmonic currents of various devices. The proposed technique is validated over a database built with a series of EMT simulations performed with various voltage and impedance conditions.