Uma abordagem estatística de processamento de sinais para detecção de ilhamento

Abstract

Unintentional islanding detection in distribution networks with distributed generation remains challenging, particularly under critical operating conditions associated with the Non-Detection Zone (NDZ), in which the generation–load mismatch may be minimal. This thesis proposes and validates a passive islanding detection method grounded on statistical signal processing and pattern recognition. The workflow is structured in two stages: (i) an offline calibration stage, in which a fixed linear projection based on event-oriented Principal Component Analysis (PCA) is estimated to compress threephase point-of-common-coupling voltages into a low-dimensional representation; next, cumulants (higher-order statistics) are extracted at nonzero lags, and features are parsimoniously selected using the Fisher Discriminant Ratio, followed by the training of decision-makers; and (ii) an online execution stage using sliding windows, in which only the previously selected lags are computed and fed into the trained decision-makers. The method was assessed on a simulated photovoltaic microgrid, considering three event classes (normal operation, islanding, and faults) and NDZ scenarios, including independent validation with unseen profiles. Under the sliding-window regime, the selected MLP classifier achieved an overall test accuracy of 98,52%, with high islanding selectivity (precision of 98,49% and sensitivity of 97,15%) and no relevant false alarms in the normal class. As a lower-complexity linear reference, a Least Squares (LS) decision-maker was also fitted in the same feature space; although it presented high precision, its overall performance and, most notably, islanding sensitivity were lower under the sliding-window regime (accuracy of 79,59% and islanding sensitivity of 66,38% in test). In the NDZ validation, the decision at the latest subwindow (worst-case delay, approximately four cycles) maintained a near-ideal performance for islanding (precision and sensitivity of 100%) and remained consistent with typical interconnection normative limits for both adopted decision-makers. Additionally, the sweep of additive white Gaussian noise from 70 to 30 dB indicated low sensitivity of the feature vector, preserving high detection rates, especially in the most informative subwindows. It is concluded that combining a fixed, islanding-oriented projection with cumulants at selected lags yields a robust, low-dimensional detector with a decision time compatible with interconnection requirements, with strong potential for practical implementation in protection functions for distribution networks.