This study proposes several high precision selective harmonics compensation schemes for an active power filter. Harmonic currents are identified and on-line tracked by novel Adaline-based architectures which work in different reference-frames resulting from specific currents or powers decompositions. Adalines are linear and adaptive neural networks which present an appropriate structure to fit and learn a weighted sum of terms. Sinusoidal signals with a frequency multiple of the fundamental frequency are synthesized and used as inputs. Therefore, the amplitude of each harmonic term can be extracted separately from the Adaline weights adjusted with a recursive LMS (Least Mean Squares) algorithm. A first method is based on the modified instantaneous powers, a second method optimizes the active currents, and a third method relies on estimated fundamental currents synchronized with the direct voltage components. By tracking the fluctuating harmonic terms, the Adalines learning process allows the compensation schemes to be well suited for on-line adaptive compensation. Digital implementations of the identification schemes are performed and their effectiveness is verified by experiments.
Harmonics Identification with Artificial Neural Networks: Application to Active Power Filtering