Desenvolvimento de base de dados e aplicação do método de Prony para extração de características e classificação de cargas elétricas

Abstract

In face of the growing demand for energy worldwide, researchers have been developing different energy management and conservation systems. Techniques such as non-intrusive load monitoring can assist in a more detailed assessment of consumption, allowing the application of energy efficiency measures, reducing the consumption of certain equipment with a greater impact on global energy consumption. In this type of approach, four main phases are considered: data collection, event detection, feature extraction, and load identification. In this context, this work proposes two main contributions: the development of a simulated database based on residential loads and the application of the Prony’s mathematical method to extract features. Regarding the database, loads are modeled in a software that allows electrical transients modeling, seeking to improve the main limitations in existing databases in the literature, such as: precise identification of events (at sample level), balanced classes, and insertion of noise and harmonics in the waveforms. As a result, a database with six sub-bases is obtained with: an ideal scenario; the presence of leakage inductance; harmonics in the network; and different noise levels. Then, using the extraction methods and load classification from the literature, it is possible to analyze whether the presence of non-idealities directly interferes on the performance of the feature extractor and load identification. In the development of the feature extractor using the Prony’s method, five different mathematical solutions are compared: polynomial, least squares, total least squares, matrix pencil and based on infinite response filters. By analyzing the classification results using this method with different classifiers (k-Nearest Neighbors, Decision Trees, Ensemble, Linear Discriminant Analysis, and Support Vector Machine), the results reached values above 90% accuracy in the vast majority of evaluated cases. It is worth mentioning that this analysis was carried out with different public databases, in addition to the database proposed here, demonstrating the feasibility of the practical use of the proposed method in the context of load classification through non-intrusive monitoring.