Machine Learning-Based Adaline Neural PQ Strategy for a Photovoltaic Integrated Shunt Active Power Filter

This paper introduces novel techniques based on Machine Learning (ML) algorithms for a Photovoltaic integrated Shunt Active Power Filter performance improvement. The first goal is to design an efficient maximum power point tracking MPPT strategy in order to harness the largest amount of energy possible. Thereby, a new hybrid Support Vector Machine Regression Perturb and Observe (SVM regression-P&O) algorithm is proposed. The SVM block improves the tracking speed by predicting an initial duty cycle, whereas a small fixed-step P&O algorithm ensures a high MPPT accuracy. The second purpose is to upgrade harmonics detection by exploiting the characteristics of intelligent learning of Adaline combined with ML algorithm. Therefore, a novel SVM regression-Adaline PQ strategy is designed. The SVM block generates the predicted initial weights of Adaline, thus ensuring fast identification of the DC active power component. In addition, the ability of this design to work with a small learning rate parameter allows an accurate harmonics extraction in contrast with the Adaptive Adaline technique where the performances are highly dependent on the chosen learning rate parameter. A comparative analysis of various ML models are carried out in order to get the best output prediction for each SVM regression block. Simulations have been performed to confirm the supremacy of the new strategies over intelligent and classical techniques. Finding exhibits a significant decrease of PV energy losses (up to 99%), a minor overshoot with an impressively decrease of the harmonics extraction’s response time (up to 98.8%), and a PVSAPF power quality enhancement under online intermittent weather conditions and variable nonlinear load.