Prototype implementation of an islanding detection relay based on pattern classification of current and voltage transients

A prototype islanding detection relay, which uses decision tree classifier to categorize the transient generating events as ‘islanding’ or ‘non-islanding’ was implemented and tested. It consisted of two basic stages of signal processing to extract the required feature vectors for the classification. The first stage involved signal filtering and in the second stage signals were processed by rectifying, summing, and lowpass filtering to get the energy content in the three phases during a selected time-frame. A simple radial medium voltage distribution system having a single distributed generator was simulated in PSCAD/EMTDC to obtain the transient waveforms, and the performance of the relay was tested with signals generated using an RTP real-time waveform playback instrument. The experimental results showed high accuracy in detecting islanding events within a response time of less than two cycles. The proposed relay can be implemented with a very low cost and is simple in construction. Therefore, this could be an acceptable low cost substitute for the expensive and complex transfer trip schemes, which are commonly in use with the distributed generation. DOI: http://dx.doi.org/10.4038/jnsfsr.v42i1.6675 J.Natn.Sci.Foundation Sri Lanka 2014 42 (1): 03-15


INTRODUCTION
The term distributed generation (DG) is referred to the small-scale generators connected to medium or low voltage power systems.They are not usually centrally planned or dispatched.DGs change the nature of the distribution system from passive to active, giving rise to certain technical issues (Jenkins et al., 2000) such as power-islanding, which is defined as 'a condition in which a portion of the utility system having both loads and distributed energy resources (DER) remains energized while isolated from the remainder of the utility system' (IEEE, 2003).Unintentional islanding results in several safety and power quality issues including abnormal variations in frequency and voltage, possibility of creating ungrounded systems, and potential safety hazards for repair crews.
In formulating guidelines and/or rules for interconnecting DGs in the distribution systems, the most common practice is to adopt the IEEE standard 1547 (IEEE, 2003).These operational practice standards have been developed with an underlying assumption that the power system is radial.In the past, the power available from DGs was not essential to support the load on the system.Hence, the standards suggest to disconnect the DGs for any 'significant' fault (a fault that could require the protection to be provoked) on the distribution system and thus, maintaining the radial nature of the system.Consequently, the islanding detection relays are designed to avoid power islanding, and these relays are referred to as anti-islanding relays (Xu et al., 2004;Mahat et al., 2008).
However, there is a change in the approach towards islanding with the initiatives on microgrids (Chowdhury et al., 2009;Lidula & Rajapakse, 2011).The performance of the power-islanding detection methods in a microgrid setup is more demanding as the microgrid is designed to operate as a stable power island.Many powerislanding detection methods in practice would find it difficult to meet the qualities demanded by micorgrids (Katiraei et al., 2005) without the use of expensive telecommunication facilities.

Prototype implementation of an islanding detection relay based on pattern classification of current and voltage transients
N.W.A. Lidula1* , A.D. Rajapakse2 , Jean-Paul Pham 2 and N. Denboer 3Different islanding detection methods proposed in the literature can be mainly categorized into three groups: passive, active and telecommunication based methods (Xu et al., 2004;Mahat et al., 2008).The passive methods make decisions based on the measured electrical quantities such as voltage and frequency (Kim & Hwang, 2000;Walling & Miller, 2002;Funabashi et al., 2003;Freitas et al., 2005a,b).In the active methods, disturbances are injected into the network and islands are detected based on the system responses to the injected disturbance (Redfern et al., 1993 (Redfern et al., 1993;Xu et al., 2004;Etxegarai et al., 2011).Passive islanding detection methods are preferred and widely used over the active and telecommunication based methods due to considerations such as the cost, simplicity and power quality.
In a situation where the intentional islanded operation is considered, there is a higher chance for the load and the generation of the island to be approximately balanced.Some passive islanding detection methods may fail to detect, or take longer time to detect the islands when such power balance exists.On the other hand, with low security levels observed in available passive anti-islanding protection, they cannot be made overly sensitive as that could lead to some unexpected control actions in a microgrid type of setup (Katiraei et al., 2005).Consequently, looking for faster and more reliable power-islanding detection methods, transfer trip schemes, which come under telecommunication based methods are emerging (Etxegarai et al., 2011).There are some recent research carried out using novel technologies such as hybrid methods, wavelets and/or classification techniques like neural networks and decision trees ( A hybrid method, which uses an average rate of voltage change to initiate a real power shift that changes the real power of DG, is proposed in Mahat et al. (2009).The proposed method has shown higher accuracy, but the islanding detection time varies from 5 to 25 cycles depending on the power imbalance.In El-Arroudi et al. (2007), the use of decision tree (DT) classifier is investigated for islanding detection.However, the proposed methodology involves some complex set of features including total harmonic distortion of current/ voltage, gradient of the product of voltage and power factor, etc. for classification.It has a misclassification rate of 16.67 % in islanding detection, which gives only 83.33 % islanding detection accuracy.The use of artificial neural network (ANN) for islanding detection has been investigated by Salman et al. (2006)

TRANSIENT BASED ISLANDING DETECTION
The transients generated during an islanding event contain unique signatures.The hypothesis that these unique signatures contained in transient signals can be used to distinguish islanding events from the other disturbances with the help of a suitable classifier, was proved in Lidula et al. (2010).It has been shown that more than 98 % overall classification accuracy can be achieved with 95 % confidence.However, hardware implementation of the above islanding detection technique requires sampling of input signal at 10 kHz and processing that data in real time.This needs high-end digital signal processors (DSP).Such expensive hardware may not be feasible for very small-scale distributed generators.This paper proposes a novel, simple and cost effective approach for classifying islanding events based on the hypothesis proved in Lidula et al. (2010) essentially by designing a novel simple feature extraction approach.
Figure 1 illustrates the block diagram of the proposed islanding detection approach, which is based on the embedded signatures of transient signals.Therefore, the measured signals need to be pre-processed to extract the required features that assist the process of classification using the decision tree (DT).Instead of using digital technologies, the feature extraction process was designed using simple analog techniques.The use of analog systems not only provides a low cost on implementation, but also gives a low computational cost compared to using feature extraction techniques such as discrete wavelet transform or S-transform in analysing transient signals.
In the proposed feature extraction method, the measured three-phase voltages (V a , V b , V c ) and currents (I a , I b , I c ) at the DG terminal were filtered into three different frequency bands (BP1: 2500 -5000 Hz, BP3: 650-1250 Hz and BP4: 312.5-650Hz) using band-pass filters as shown in Figure 1 The following subsections briefly discuss the theory relevant to each component used in the implementation.

Band-pass filters
Fourth-order Butterworth band-pass filters (BPF) were applied in this study.The filters were designed by cascading two second-order BPF that use the multiple feedback (MFB) topology.The MFB topology allows the gain, bandwidth, and mid-frequency to be adjusted independently of one another.The fourth-order MFB band-pass filter topology is shown in Figure 2 (Kugelstadt, 2002).The general equation for the transfer function of a second order Butterworth band-pass filter is given by equation ( 1).
With the cascading of two 2 nd order BPFs to get a 4 th order BPF, the transfer function thus, can be given by equation ( 2).
where, α is a scaling factor determined through successive approximation of equation (3).Q i and A mi are the quality factor and gain, respectively of both individual 2 nd order March 2014 Journal of the National Science Foundation of Sri Lanka 42 (1) filters.Once α is found, the values of Q i and A mi can be found from equations ( 4) and ( 5), respectively.Once Q i and A mi are known, the resistor values are calculated using relationships shown in Table 1, for a pre-defined value of capacitance (10 nF).

Rectifiers
After filtering out the waveforms into different frequency bands, the features were formed by calculating the energy content of each filtered signal in a selected time window.The first stage of the energy extraction process is the fullwave rectifier.

Adders
In order to create a three-phase energy value, the rectified three-phase inputs were summed together using an inverting adder circuit shown in Figure 4.The output of this circuit is given by equation ( 6).Additionally, the three phase adder circuit allows amplifying the signals by adjusting the ratio of R f /R.Thus, the inverting topology effectively supports the cancellation of the negative gain of the next step of the energy extraction process: the low-pass filter.

Low-pass filters
The final stage of the energy extraction process is a second-order Butterworth low-pass filter.The low-pass filter performs the integration of the waveform over a definite time window (0.01 s).The circuit diagram of the second-order low-pass MFB topology is shown in Figure 5.The general transfer function of a low-pass filter is given by equation ( 7), which gives equation ( 8) for the MFB low-pass filter shown in Figure 5 (Kugelstadt, 2002).
...( 7) For a given cut-off frequency, f c , C 1 and C 2 the resistor values can be calculated using the relationships given in Table 2.In selecting C 2 , it is required to satisfy equation ( 9) for R 2 to be a real value.CART algorithm generates decision trees based on a splitting rule.The basic idea of the splitting rule is to choose a split (attribute to test) among all the possible splits at each node so that the resulting child nodes are the 'purest'.In the CART algorithm, each split depends on the value of only one predictor variable (input feature).It carries out a brute force search through all possible splits to find the particular split.The 'Gini rule,' which searches for the largest class in the learning dataset to isolate it from the rest of the data was selected as the splitting rule in this study.For a particular split s at node t, the Gini impurity measure i(t) is defined as: where, t is the node number, Y j = class ( j = 1,2, .... J) and p(Y j | t) is the class probability distribution of the dependent variable at node t.
Once the impurity measure is calculated, then the goodness of split s, ∆i(s, t) is calculated using equation (11).
where p L is the proportion of cases at node t that go into the left child node, t L and p R is the proportion of cases at node t that go into the right child node, t R , i( t L ) is the impurity of the left child node and i(t R ) is the impurity of the right child node.
The best split point is the one that maximizes the goodness of split, ∆i(s,t) when the node is split according CART also provides a ranking of input features based on each variable's contribution to the overall tree.Variable importance is presented by scaling the values relative to the 'best performing variable', which is the variable that minimizes the Gini impurity the most at every node it is considered as a splitter (Breiman et al., 1984

Test system
The simple radial MV distribution network shown in Figure 7 was simulated using PSCAD/EMTDC power system simulation software.
The distribution network consists of five bus-bars each having lumped loads, and bus-5 has a 1 MW, 0.3 MVar synchronous generator connected via a 13.2/0.4kV transformer.All loads in the network were simulated as constant RL loads.The nominal system frequency is 60 Hz.It was intended to classify the system disturbances into two classes of events, namely 'non-islanding' and 'islanding'.The non-islanding cases include: (i) normal operation; (ii) switching of loads and (iii) temporary faults on bus-2.The islanding cases simulated include: (i) opening of breaker 'Tripgrid'; (ii) opening of breakers 'Trip' and (iii) opening of breakers 'Trip' following a fault on bus-2.
The simulations were carried out under different loading conditions by changing the system total load between 3 MW to 10 MW keeping a power factor close to 0.95.The tripping and fault times were also varied.The cases with faults were simulated for three-phase (ABC), three-phase-to-ground (ABCG), line-to-line (AB, AC, BC), and line-to-ground (AG, BG, CG) faults, where the fault resistance was also varied.A total of 864 islanding and 510 non-islanding cases were simulated and the three-phase energy values of terminal voltage and current were recorded for each simulation.
To make the data collection consistent, a 'trigger' signal was defined.It is a transient detector, which is based on the three phase energy value of the voltage waveform output through the band-pass filter in the range 312.5 − 650 Hz (defined as EV4).The selection of EV4 as the trigger was straightforward due to its quality of showing a comparatively large variation in the event of a transient.The trigger threshold was set as 0.4 by studying the dataset.Thus, the data was collected after a delay of 0.01 s, if the value of EV4 exceeds 0.4.The delay of 0.01 s allows the required energy extraction process to be completed.

Classifier modelling
The decision tree classifier was modelled based on the theory explained earlier, using the CART ® Pro V6 commercial software (Steinberg & Colla, 1995).From the total dataset, which included a total of 864 islanding and 510 non-islanding cases, 67 % was used as the data, while the remaining 33 % was used as the testing data.The testing data were extracted randomly from each category of events described above to ensure testing against all types of transient events.The optimized DT shown in Figure 8 returned the same classification accuracy as the initial DT, and the classification accuracies are given in Table 4.The overall accuracy (correct classification of both islanding and nonislanding events as a percentage of total events tested) of 100 % was achieved in training the classifier.In testing, the overall accuracy achieved was 99.83 %, while giving 100 % accuracy in classifying islanding events.
Figure 9 presents the simulation results observed at each point of signal processing from the voltage and current outputs (PT and CT outputs) measured at the DG terminal up to the classifier for a non-islanding (temporary three-phase fault with 1 Ω fault impedance at the bus-2 in Figure 7) and an islanding (opening of breaker 'Tripgrid' in Figure 7) event.Figures 9(c) -(g) are intended to illustrate the signal processing in each step under feature extraction for a non-islanding (fault) and an islanding event.
Feature extraction involves band-pass filtering, rectifying, three phase adding and low pass filtering.Only the band-pass filter outputs of phase-a are shown for the two cases (non-islanding to the left and islanding to the right) in Figures 9(c) and (d).Similarly, the rectifier output is shown only for the frequency range 312.5 < f < 650 Hz (BP4) of phase-a voltage in Figure 9(e).The adder output is shown only for the BP4 frequency range in Figure 9(f).This comparison in Figures 9(c) to (g) depicts how the signal processing supports creating distinguishable differences (specifically in magnitude) in the features (EV1, EV3, EV4 and EI4) finally shown in

Prototype implementation
A prototype of the proposed islanding detection methodology was implemented in the laboratory level to investigate its real time performance.The main advantage in this proposed method compared to Lidula & Rajapakse (2010) arises in the implementation.Higher computational cost was avoided with the elimination of the DWT from feature extraction.In this proposing methodology, the feature extraction part was separately implemented using the actual filters.The output signals of the filters were used as inputs to the DT classifier, which was implemented on a microprocessor.The details of implementation of the feature extraction process and DT classifier are described in the following sub-sections.

Feature extraction
As shown in Figure 1 and as discussed earlier, the feature extraction process begins with the band-pass filtering of the three phase voltages and currents.The estimated values for all the components of the fourthorder Butterworth band-pass filters (Figure 2) are given in Table 5.A total of 12 fourth-order band-pass filters (3 filters for each phase of voltage, and 1 filter for each phase of current) were implemented.
The next stage of feature extraction is the rectifier circuits.The brief discussion on rectifiers in a previous section shows that each rectifier circuit consists of two operational amplifiers, five resistors and two diodes.The selected values for the resistances are shown on Figure 3 (Gerstenhaber et al., 2010).
The third stage of the process is the adder circuit.The resistor values of the adder circuits (R f and R) were selected to obtain the required amplification gains for voltage and current signals (4 and 30, respectively).
The final stage of the feature extraction is low-pass filtering.As discussed earlier second-order Butterworth filters were used considering their capability in giving smooth output with minimal distortions.The component values used for the low pass filters corresponding to Figure 6 were determined as shown in Table 6.The output from the four low-pass filters represents the four energy signals that act as predictors to the classifier.

Classifier
The classification algorithm was implemented using the Microchip PIC18F46K22 microcontroller.The key features of this microcontroller are given below.The microcontroller was incorporated with a MPLAB ICD2 connector, resistors, a seven-segment display for testing and light emitting diodes for logic flags.The developed classification unit is shown in Figure 10.The required software modules for converting analog signals into digital values and displaying characters on a 4 digit seven segment display were developed.Furthermore, the subroutine for the DT classifier was directly imported from the CART software to programme the microchip.
The instantaneous energy values EV1, EV3, EV4 and EI4 resulting from the low-pass-filter outputs were fed to the microchip through the MPLAB ICD2 connector.The system first reads EV4 and compares the value to the trigger level (0.4).Once EV4 surpasses this value, a delay of 0.01 s is provided before collecting EV1, EV3, EV4 and EI4.These values are used as the input parameters for the DT classifier subroutine, which returns an integer representing the result of the classifier.Upon detection of an islanding event, an islanding flag is set to high, and the programme ends.If a non-islanding event occurs, the system waits for EV4 to decrease to a value below the trigger, and the process repeats until an island is detected.

Experimental setup
The waveform generator, real-time playback (RTP) was used to generate the necessary three phase voltage and current signals to test the classification system.The recorded waveforms obtained from PSCAD/EMTDC simulation were played back via the RTP.A photograph of the experimental setup used for prototype testing is shown in Figure 11.

RESULTS
The implemented prototype transient based islanding detection relay was tested using 26 islanding and 26 nonislanding events.

Validation of the experimental setup operation
In order to ensure correct operation of the classifier, the analog waveforms generated from the simulations were compared against the actual waveforms observed in the implemented islanding detection relay.Figure 12 shows a comparison of the simulated and measured signals observed during an islanding transient.Figure 12(c) shows that the energy values obtained from the implemented actual filters are in total agreement with the simulated values.

Classification results
The operation of the implemented classifier was tested for 52 cases, which involved 26 islanding events and 26 non-islanding events.Figure 13 shows the performance The results observed with the implemented islanding detection relay for the 52 cases tested are summarized in Table 7.It shows that 24 out of the 26 islanding cases tested were correctly classified as islanding, and 25 out of the 26 non-islanding cases were correctly classified as non-islanding.This gives an overall classification accuracy 94 %.The islanding cases that were incorrectly classified operated due to the transients originating from fault reflections.In average, the islanding detection time of the implemented relay was estimated as 14 ms.The maximum detection time observed is only 19 ms, which is less than 2 cycles.

CONCLUSION
Development of a prototype islanding detection relay transient signals was presented in this paper.The islanding detection relay uses a decision tree classifier with the features extracted through a set of tuned analog band pass filters.The proposed islanding detection technique was not only successful in giving comparatively accurate classification results, but also presented a simplified design with reduced computational and implementation cost.
In testing the implemented prototype relay, classification accuracies matched very well with the simulated results.With the implemented relay, an overall accuracy of 94 % was achieved in classifying islanding and non-islanding events correctly.For the cases tested, the average islanding detections time was only 14 ms, which is less than one cycle and the maximum detection time observed was only 19 ms, which is less than two cycles.
With its fast response, higher accuracy level, and the potential for low cost implementation, this methodology can be accepted as a promising technology to apply in microgrids, which require fast and reliable islanding detection (Katiraei et al., 2005) and could be a challenging substitute for highly reliable but expensive and complex transfer trip schemes (Etxegarai et al., 2011).
Figure 1 illustrates the block diagram of the proposed islanding detection approach, which is based on the embedded signatures of transient signals.Figure 1 depicts that islanding detection is completed in three major stages: (1) data collection, (2) feature extraction and (3) classification.At data collection, current and voltage signals are measured at the generator terminal with the

Figure 1 :
Figure 1: The proposed islanding detection method .The required frequency bands were selected based on Lidula et al. (2010) considering the DWT coefficients used, and further optimized through the process of building the optimal decision tree classifier.The three-phase outputs of current and voltage in each frequency band (i.e. in the frequency band BP1: I a 1, I b 1, I c 1 and V a 1, V b 1, V c 1) were then rectified using full-wave rectifiers (i.e.giving |I a 1|, |I b 1|, |I c 1| and |V a 1|, |V b 1|, |V c 1| in BP1) and fed into three input adders, where three phase values were summed together (i.e. in BP1, V1=|V a 1|+|V b 1|+|V c 1| and I1=|I a 1|+|I b 1|+|I c 1|).The lowpass filters were finally used to integrate the adder output in a definite time window creating the required features (i.e. in BP1: EV1 and EI1) for the decision tree classifier.

Figure 5 .
Figure 5. Second-order low-pass filter circuit(Kugelstadt, 2002) A decision tree (DT) is constructed based on the training data, and represented as a binary tree.The DT starts with the 'Root', and it contains the total training dataset.Starting from the 'Root', at each 'Internal Child Node' an attribute is tested and the data is separated to two groups based on the agreement and disagreement with the attribute.The 'Terminal Node' represents the predicted class(Breiman et al., 1984;Steinberg & Colla, 1995).Figure6shows a simple structure of a DT for a two class problem.The CART ® Pro V6 commercial software(Steinberg & Colla, 1995) was used to develop the decision tree classifier.

Figure 6 :
Figure 6: A simple model of a decision tree

Figure 8 :
Figure 8: The decision tree classifier trained for detecting power islands

Figure 9 (
Figure9(g) for the non-islanding and islanding events, which are the inputs to the DT classifier.

Figure 9 :
Figure 9: The signal processing process in the proposed relay for a non-islanding and an islanding event: (a) terminal voltage, (b) terminal current, (c) band-pass filters' outputs of phase-a voltages, (d) 312.5 − 650 Hz band-pass filter (BP4) output of phase-a currents, (e) rectifier outputs of phase-a voltage in the range 312.5 − 650 Hz (BP4) (f) 3-phase adder output of voltages in the range 312.5 − 650 Hz band (BP4) and (g) low-pass filter outputs

Figure 12 :
Figure 12: Comparison of simulated and measured signals (a) phase a-voltage, (b) phase-a current and (c) energy waveforms

Table 1 :
Relationships : the resistances of the 4 th order band-pass filter

Table 2 :
Relationships: the resistances of the second order low-pass filter

Table 3 :
Variable importance for initial DT with 8 variables each variable in the classification process as a percentage of the 'best performing variable' which is shown in Table3.According to Table3, relevancy of the variables EI1, EI2, EI3 and EV2 are very low, and they were omitted in

Table 4 :
DT classification accuracies through CART software

Table 6 :
Standard capacitor and resistor values for the low-pass filters

Table 5 :
Standard resistor values for the band-pass filters

Table 7 :
Experimental accuracy results