Aplicação de algoritmos de aprendizado supervisionado na identificação de padrões elétricos de corrente e tensão

Abstract

Non-Intrusive Load Monitoring (NILM) has established itself as a key technology for energy efficiency in smart grids, enabling consumption disaggregation without the need for multiple sensors. This work proposes and validates a computational methodology for the automatic identification of electrical loads, grounded in Data Science and Machine Learning techniques. Unlike approaches that rely on data collection in controlled environments, this study utilizes the unification of two heterogeneous public datasets (PLAID and COOLL), ensuring greater robustness and generalization capability against noise. The feature extraction strategy adopts a hybrid approach, combining spectral analysis via Fast Fourier Transform (FFT) and timefrequency decomposition via Discrete Wavelet Transform (DWT) with statistical descriptors in the time domain, such as Crest Factor and Kurtosis. Three supervised classification algorithms were comparatively evaluated: k-Nearest Neighbors (k-NN), Support Vector Machines (SVM), and Random Forest. The experimental results, validated on an independent test set, demonstrated the superiority of the Random Forest model, which achieved an F1-Score of 0.95, overcoming the class imbalance limitations observed in the other models. Additionally, the computational cost analysis confirmed the theoretical technical feasibility of implementing the algorithm in low-cost embedded systems (Edge Computing), validating the solution for future applications in IoT devices.