Islanding detection technique based on relevance vector machine

Abstract

This study present a new islanding detection technique based on relevance vector machine (RVM) containing various types of distributed generations (DGs). The proposed scheme is based on utilising negative sequence component of current (I2), acquired at the terminal of the target DG. Various islanding and non‐islanding events with variable real and reactive power, change in network topology and diverse X/R ratio have been generated by modelling IEEE 34 bus system in real time digital simulator (RTDS®/RSCAD) environment. The samples of I2 for one cycle duration from the inception of islanding/non‐islanding events are given as input to the proposed RVM classifier. The proposed classifier is able to discriminate between islanding and non‐islanding events with an accuracy of 98.62% by utilising only 40% of the total dataset in training. Moreover, it is capable to classify islanding condition rapidly and correctly even with perfect power balance situation. Furthermore, it remains immune to change in network configuration/entirely different network and X/R ratio of various types of DGs. Comparative analysis of the proposed scheme in terms of accuracy, testing time and number of relevance vectors (RVs) with the existing schemes shows its superiority in discriminating islanding situation with non‐islanding events.

Auto‐correlation‐based Islanding detection technique verified through hardware‐in‐loop testing

“…Therefore, the overall Islanding detection time from the inception of the Islanding event would be around 60 ms which is equivalent to three cycles. Table 7 shows the comparative assessment of the proposed scheme in terms of the percentage value of NDZ and detection time with the techniques based on ROCOF [11], oscillatory frequency [12], overvoltage/undervoltage (OV/UV), over frequency/under frequency (OF/UF) [13], inverse hyperbolic secant function (applied for acquired voltage signals) [16], time-frequency (TF) transform [17], wavelet transform (WT) [18][19][20], Hilbert-Hung transform (HT) [21], SVM [22,23], RVM [25], artificial neural network (ANN) [26,27], adaptive ensemble classifier (AEC) [28], Data mining [29,30], random forest (RF) [32], and principle component analysis (PCA) [24].…”

“…Furthermore, the hardware implementation of said techniques is also complex [17][18][19][20][21]. Subsequently, support vector machine (SVM), relevance vector machine (RVM), random forest, neural network, adaptive ensemble classifier, data mining, and principle component analysis-based approaches have been discussed in [22][23][24][25][26][27][28][29][30][31][32]. Even though these approaches give good results, the requirement of a vast number of input patterns for training, complexity in training procedure and large errors for unobserved pattern/dataset make the above techniques less popular.…”

Synchrophasor based islanding detection for microgrids using moving window principal component analysis and extended mathematical morphology

“…During balanced generation-load conditions on the island, variation in voltage, current and frequency signals are negligible. The use of PCA alone results in the non-detection of IE during the perfect power balance condition [32]. However, signal processing methods can reduce NDZ significantly by capturing the tiniest variation and sensing hidden features in the measured signals [33][34][35][36][37].…”

Passive method for islanding detection using variational mode decomposition

“…On the other side, passive IDTs can be classified into time and frequency domain IDTs [13–18]. In passive IDTs, under/over‐voltage and frequency relays and rate of change of frequency relay are the most used schemes for islanding protection [19].…”