Application of Ensemble Learning with Mean Shift Clustering for Output Profile Classification and Anomaly Detection in Energy Production of Grid-Tied Photovoltaic System

Abstract

Fault detection and monitoring system in photovoltaic (PV) energy management system is important in achieving its optimal performance. An effective diagnostic system involves correct analysis of electrical parameters of a PV array on a given weather condition. In the study, mean-shift clustering was applied for pre-classification and anomaly detection of time-series data of electrical parameters from grid-tied inverter, and solar-irradiance. Classification and anomaly detection applied is based in ensemble learning, where its base learners are based from multilayer perceptron. A stacking ensemble is used in classification of energy production profile while bagging ensemble is used detecting anomalous trend in time-series data. A stacking ensemble got a highest accuracy value of 94% compared to single classifiers which have accuracy value of 85.25%, 84.14%, and 63.4%, respectively. The bagging ensemble autoencoders have the lowest mean squared error during model reconstruction compared to single autoencoder. It has a fair performance in classifying anomaly points from normal datapoints, having an AUC value of 0.795 and F1-score of 0.71, given that the hyperparameter is 0.5. Overall, ensemble learners improve the performance in classification and detection tasks.