ML and DWT for Power Quality classification (Master Thesis)

Machine Learning Applications for a Robust Classifier of Advanced Power Quality Disturbances in Smart Grid.

This project is associated with a master's thesis entitled "Machine Learning Applications for a Robust Classifier of Advanced Power Quality Disturbances in Smart Grid", by Gabriel C. S. Almeida, at the Federal University of Itajubá. Here are the MATLAB code in detail.

The insertion of new devices, increased data flow, intermittent generation and massive computerization have considerably increased current electrical systems’ complexity. This increase resulted in necessary changes, such as the need for more intelligent electrical networks to adapt to this different reality. The generation of Artificial Intelligence (AI) technology represented by Big Data, Machine Learning (ML), Deep Learning (DL) and Pattern Recognition represents a new era in society and global development based on information and knowledge. With the recent Smart Grids (SG), the use of techniques that use this type of intelligence will be even more necessary. This dissertation investigates the use of advanced signal processing and ML algorithms to create a Robust Classifier of Advanced Power Quality Disturbances in SG. For this purpose, known models of PQ disturbances were generated with random elements to approach real applications. From these models, thousands of signals were generated with the performance of these disturbances. Signal processing techniques using Discrete Wavelet Transform (DWT) were used to extract the signal’s main characteristics. This research aims to use ML algorithms to classify these data according to their respective features. ML algorithms were trained, validated, and tested. Also, the accuracy and confusion matrix were analyzed, relating the logic behind the results. The stages of data generation, feature extraction and feature selection were performed in the MATLAB software. The Classification Learner toolbox was used for training, validation and testing the 27 different ML algorithms and assess each performance. All stages of the work were previously idealized, enabling their correct development and execution. The results show that the Cubic Support Vector Machine (SVM) classifier achieved the maximum accuracy of all algorithms, indicating the effectiveness of the proposed method for classification. Considerations about the results were interpreted as explaining the performance of each technique, its relations and their respective justifications.