Optimal Machine-Learning-Based Controller for Shunt Active Power Filter by Auto Machine Learning

• Published in: IEEE journal of emerging and selected topics in power electronics Vol. 11; no. 3; pp. 3435 - 3444
• Main Authors: Muneer, V., Biju, G. M., Bhattacharya, Avik
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.06.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Active filters; Algorithms; Artificial neural network (ANN); Artificial neural networks; AutoKeras; automated machine learning (AutoML); AutoSklearn; Bayesian optimization (BO); Controllers; Data models; Decision trees; Deep learning; Design optimization; Gaussian process; Load modeling; Machine learning; machine learning (ML); Mathematical models; Multilayer perceptrons; network morphism; neural architecture search (NAS); Neural networks; Optimization; power quality (PQ) problems; Predictive models; Preprocessing; Recurrent neural networks; shunt active power filter (SAPF); Support vector machines; Voltage control
• ISSN: 2168-6777; 2168-6785
• DOI: 10.1109/JESTPE.2023.3244605

This article proposes a machine-learning (ML)-based controller for shunt active power filter (SAPF). ML design had always been suboptimal as human experts designed them, even though ML models are computationally simpler than traditional methods. This drawback is also addressed by optimizing all the design processes, including data preprocessing, model selection, and model hyperparameters. Automated ML (AutoML) algorithms are used to find the optimum pipeline for an ML model from a search space of 110 hyperparameters (more than 15 ML models, 14 feature preprocessing methods, and four data preprocessing methods). It has both classical ML methods and deep-learning models in its pool, such as decision trees, support vector machines, ADALINE, multilayer perceptron (MLP), recurrent neural networks (RNNs), convolutional neural networks, Gaussian process (GP), and so on. Optimization is done with Bayesian optimization (BO)- and network morphism-based neural architecture search (NAS). This work proposes a method to develop a general artificial neural network (ANN) model independent of system parameters. This universal ML controller produces accurate current reference generation and faster estimation results compared to the conventional control scheme. The proposed techniques are simulated with MATLAB and Python. The proposed system was also verified using an experimental prototype developed in the laboratory.