Fault Detection in Microgrids Using Combined Classification Algorithms and Feature Selection Methods

Due to fault current variations over a wide range, protection strategies relying on high fault currents in microgrids are a big challenge. This paper proposes a method for fault detection in microgrids using data mining patterns and classification algorithms. For this reason, several short circuit fault and no-fault cases (i.e. load switching, motor starting and transformer energization) are generated and one cycle of the voltage and current signals is preprocessed by wavelet packet transform (WPT). The main features of voltage and current signals are extracted using detailed coefficients of the WPT. For discriminating faults from no-fault events, two different classifiers (i.e. random forest (RF) and K-nearest neighbors (K-NN)) are utilized. To improve the classifiers accuracy or reduce data storage requirements of relays, two filter based feature selection methods are applied on feature vector for choosing the most relevant features. For evaluating the performance of the complete and reduced feature vectors, the standard IEC microgrid is simulated for both islanded and grid connected modes of operation with meshed and radial structures. Test results show the effectiveness of the proposed method for fault detection.