Learning and adaptive techniques for harmonics compensation in power supply networks

This paper compares different variants of the least mean squares (LMS) algorithm. The objective consists in finding the best compromise between on-line learning and computational costs. Indeed, an algorithm with low computational complexity for updating Adalines weights is required for a real-time implementation of a modular neural Active Power Filter (APF). This filtering scheme is inserted in an electric distribution system to identify and compensate for harmonic distortions. Adaline learning schemes are used in two neural APF frameworks. The first one is a neural approach of the Instantaneous Power Theory (IPT) where the instantaneous powers are decomposed in a linear manner and are learned on-line with Adalines. The second one is a neural diphase currents method based on the DQ-currents which are linearly decomposed and learned with Adalines. The overall complexity of the neural frameworks is evaluated in terms of basic operators such as adders, multipliers, and signum functions. Simulation and experimental results demonstrate the applicability of neural approaches for the control of APF frameworks for power quality improvement. The complexity of the neural APF frameworks are equivalent than methods based on the conventional Instantaneous Power Theory (IPT), while their performances are superior.