Estimador seletivo do conteúdo harmônico de tensão e corrente baseado em rede neural profunda

Abstract

A common issue when non-linear loads are present in electric power distribution systems is voltage and/or current harmonic distortion. This power quality (PQ) problem can be mitigated, but first the harmonic components need to be identified or estimated. The main goal of this work is to develop a method for the selective estimation of amplitudes and phase shifts of the 3, 5, 7 and 9 harmonics from quarter-cycle samples of unknown waveforms, based on deep neural networks (DNN). A sample set of quarter-cycle current waveforms was generated for DNN training, validation and testing. An grid search for parameters was used together with cross validation to define the configuration of the computational model, which resulted in the DNN configuration used in this work. Other regression methods were compared to DNN to justify its use, showing that the neural network method is capable of achieving lower errors in relation to the other methods tested. Test results show that it is possible to achieve total harmonic distortion (THD) attenuation by means of an ideal active power filter (APF). This APF receives the reference current generated through the amplitudes and phase angles estimated by the proposed DNN. An THD attenuation was achieved from 18.54% to 0.81% in the case study carried out with ideal APF. The method validation took place through computer simulations that demonstrated the ability to selectively estimate harmonics through a current signal with a quarter-cycle sampling. The main contribution of this work is the proposal of a selective harmonic estimation method that can be used, for example, in an APF, for harmonic distortion attenuation, or in PQ monitoring and control applications.