Deep Learning-based control of active power filters using LSTM and GRU networks for harmonic and frequency estimation

Abstract

Deep Learning (DL) techniques provide a powerful tool enhancing the learning capabilities of the neural networks (NN), and are increasingly applied in the field of electric power systems. In particular, the long short-term memory (LSTM) and the gated recurrent unit (GRU) networks allow improvements on signal processing. The relevance of suppressing electrical disturbances justifies the efforts to apply new control algorithms to the active power filters (APF). Despite the existence of many control techniques, the NN-based proposals generally present significant shortcomings. Therefore, in this work, a new neural controller is presented for further improvement, using previously trained NNs, without need of adaptive algorithms. The generation of the three-phase APF reference currents is based on LSTM and GRU networks, that extract the full necessary information from currents and voltages, thus avoiding the need of an additional phase synchronization control. It is a novel proposal comprising FCE (fundamental Fourier coefficients estimation) and FE (frequency estimation) along with a simple computation process, for harmonic distortion and reactive power compensation. It has been tested with many practical loads and conditions through simulation and experimental platforms. Its general high performance confirms a substantial progress compared to other NN controllers, and it could be an alternative to other techniques.