Comparative Study of Different Approaches for Islanding Detection of Distributed Generation Systems

Abstract

The issue of unintentional islanding in grid interconnection still remains a challenge in grid-connected, Distributed Generation System (DGS). This study discusses the general overview of popular islanding detection methods. Because of the various Distributed Generation (DG) types, their sizes connected to the distribution networks, and, due to the concern associated with out-of-phase reclosing, anti-islanding continues to be an issue, where no clear solution exists. The passive islanding detection technique is the simplest method to detect the islanding condition which compares the existing parameters of the system having some threshold values. This study first presents an auto-ground approach, which is based on the application of three-phase, short-circuit to the islanded distribution system just to reclose and re-energize the system. After that, the data mining-decision tree algorithm is implemented on a typical distribution system with multiple DGs. The results from both of the techniques have been accomplished and verified by determining the Non-Detection Zone (NDZ), which satisfies the IEEE standards of 2 s execution time. From the analysis, it is concluded that the decision tree approach is effective and highly accurate to detect the islanding state in DGs. These simulations in detail compare the old and new methods, clearly highlighting the progress in the field of islanding detection.

1. Introduction

In the present power system with distributed generators (DGs), there are some issues that are not resolved yet. One of these issues involves a fast and very accurate detection of islanding [1,2]. The number of DGs introduced into electricity distribution systems is increasing day by day, and it is the most challenging concept in modern power system scenarios, since DG islanding can cause degraded proficiency, excellence, reliability and quality of supply [3,4]. The concept of Distributed Energy Resources (DER) is moving from being a local issue towards a system issue [5]. With the connection of DGs, the power system becomes more complicated [6]. With the introduction of the reference power transition link, the grid connection process is further ensured to be smooth [7,8]. The main difficulties that researchers encounter are not the lack of data, but the efficiency of using complex electrical quantities [9]. The issue of unintentional islanding in DGs grid interconnection still remains a huge challenge in grid-connected systems, since the system may not succeed at activating the protective devices during the islanded condition [10,11]. This results in the issues related to the power quality, protection, reverse power flow and system stability [12]. Hence, the condition of islanding must be detected and cleared as fast as possible as per the recommendation of grid codes [13]. The maximum time to respond to the islanding detection is recommended to be 2 s by IEEE [14]. The islanding protection devices being used in the future have to reflect this. Safety measures must be taken for correct operation of these new DGs so as to prevent uncontrolled and undesired power injection into the grid and possible damages to equipment and personnel [15]. Thus, a thorough study is needed in solving this issue. Hence, anti-islanding with fast response time is essential for a DG connected grid system [10,16]. To overcome the challenges from islanding DGs, researchers proposed a numerous model that deals with the consequences of intentional islanding, and clears it as fast as possible [11,12,17,18,19,20,21,22,23,24,25,26,27,28,29]. Different organizations such as the IEEE, IEA and IEC also set the standards and were updated regularly to emphasize the importance of islanding detection in a DG connected grid, so that a DG connected system will operate smoothly [14,30,31,32].

In this study, the comparative study of different islanding detection is discussed based on the literature review. This paper is intended to study the comparative analysis between two techniques: one traditional auto grounding method, and a relatively new data-mining-decision-tree technique. This study first introduces the issues of islanding detection techniques and its requirements. Section 2 presents the literature review of different topics associated with the concept of islanding and anti-islanding methods, and their comparisons. In Section 3, applied methodologies and approaches are explained. Section 4 presents the research outcome and their discussion. Finally, in Section 5, conclusions are drawn.

2. Literature Review

2.1. Distributed Generation

Distributed generation (DG) is the system that indicates the energy sources that provide reliable electricity to the end users. These technologies are site-specific and referred to small renewable energy sources such as solar, wind, micro-hydro. It can be at a local level or end-user level. The local level DG is the renewable energy systems that electrify a community, whereas the end-user level includes the technologies adopted by the single consumer with similar characteristics of the system. The energy storage system plays a vital complementary role in DG. Although being the decentralization and the smaller size, DG and the energy storage system are highly preferred for low energy costs with high reliability, and a more efficient, cost-effective and eco-friendly nature [33,34,35,36,37,38,39,40].

Figure 1 presents an example of a DGS containing micro-hydro power plant, solar PV, wind turbine, diesel generator as the energy-producing units, the battery bank as a storage device, load, and finally grid as an optional choice. The voltage level of AC and DC bus bars may be the same or different, but in general the DC level is low and controlled through the converter. A circuit breaker (CB) is placed between the grid and the AC bus system. Here, if the CB is open, the energy system, excluding the grid system, can be considered as IES, and, if the status of CB is closed, the system will be grid-connected. In Figure 1, the system can be operated in either an isolated or grid-connected mode. The flow of the current in the CB section may be unidirectional or bidirectional, depending on the nature of the connection and the local policy of energy trading.

Conventionally, the energy systems are of two types; isolated and grid connected system. The isolated energy system (IES) is a collection of energy generator, load and the controlling components that are designed to fulfill the energy demand of a community or a system in isolated form. There is no connection of the national grid line to get power from any reliable service. Generally, IES seems to be a promising option for electrifying remote locations where grid extension is not feasible or economical. The integration of battery banks as a means of energy storage with different renewable energy systems can enhance the system reliability and its overall performance. Therefore, appropriate choices of generator sizes and the battery bank capacity are critical to the success of such renewable-based, isolated power systems. However, there are numerous challenges in renewable energy based IES such as high-cost, poor reliability and power quality, low load factor, periodic nature of renewable sources, problems in maintenance and monitoring activities [41]. Because of the periodic nature and dependency on weather factors, the characteristics of the energy generated from renewable energy sources (RES) like solar and wind are unreliable, which can be improved by using sufficient storage devices or by interconnecting multiple energy resources type, called optimal hybridization [42,43]. An optimized hybrid energy system (HES) is quite a promising technique on behalf of cost, power quality and reliability able to provide the electricity to the area where expansion of the national utility’s grid system is quite difficult and expensive [44,45,46]. It can be made more efficient and cost-effective by reducing the disadvantages associated with these technologies [41,47]. However, there are other problems as well, such as: low load factor, low diversity factor, low reliability, protection, etc. in the isolated energy system. To resolve these problems, the mini-grid concept may be a good option. It is proven that the mini-grid concept contributes to the enhancement of social impact by improving the local governance structure via direct participation of the local community, since it is the best option to electrify the isolated rural communities [45,48]. On the other hand, a grid-connected system is the energy system containing an energy generator, as well as load components that are interconnected with a utility.

2.2. Islanding Detection Methods

2.2.1. Passive Method

Passive method is based on continuous monitoring of the local site system parameters like change in voltage, frequency, and harmonic distortion as these parameters vary during the islanding condition initiation [49]. These variations are directly influenced by the power mismatch between the supply and the loads. Due to passive methods of islanding detection being dependent on the changes of the aforementioned parameters, the efficiency is thus dependent on the threshold set [3,50]. Setting a lower threshold results in false islanding detection as non-islanding faulty conditions of the power system may cause system parameter changes too. The higher setting of the threshold results in the occurrence of islanding condition being undetected as system parameter changes during islanding will not cross the set thresholds [51]. Though cheap on implementation, inability of the method to detect islanding during balanced islanding and large non-detectable zone (NDZ) renders this method unreliable to be used alone [52]. The rate of change of voltage/frequency relies on the fact that power imbalance causes transients in an islanded system and causes the voltage and frequency to vary. The over/under voltage and frequency, however, deal with the thresholds of the voltage and frequency, respectively, to detect an islanded system. The Harmonics Detection Method monitors the total harmonic distortion in the grid, while the phase monitoring method constantly detects any sudden changes in the output current of the inverter and its terminal voltage to detect the islanding condition [53]. Explanations to some of the passive methods are:

A. Rate of Change of Frequency:

This method uses the equation:

$$ROCOF= \frac{df}{dt}=\frac{∆\mathrm{P}}{2H} f.$$

(1)

This method’s major drawback is that it cannot detect the islanding condition until the active power imbalance is higher than 15%. It detects the load and generation mismatch, which in turn affect the frequency. This change in frequency is then monitored to trip the relay, thereby disconnecting the DG and the power grid [54].

B. Rate of Change of Voltage:

This method involves calculating of rate of change of voltage given by:

$$ROCOV= \frac{dV}{dt}.$$

(2)

It is based on the fact that the reactive power imbalance causes a transient change and voltage starts to vary dynamically. Both ROCOF and ROCOV were tested with the conclusion that they can detect islanding condition within 0.125 s [55].

C. Over/Under Voltage and/or Frequency:

This method of detection of the islanding problem is the simplest of the available methods where the voltage and/or frequency is continuously measured and when the voltage or frequency is measured to be outside the predetermined acceptable limits, the connection between the DG and the power grid is disrupted. In the case that the power rating of the DG is less than the power rating of the load, the mismatch of the power will be compensated by the grid. However, when the islanding occurs, there will not be any power supply from the grid, so in this case the voltage and frequency will deviate. That is, in this case, when there is islanding effect, the voltage at the Point of Common Coupling (PCC) should increase until active power P (DG) = Real Power of the load P (load). The Non-Detection Zone (NDZ) refers to the zone of within which the islanding condition is not detected by the islanding detection method.

The frequency should also change such that the reactive power Q (DG) is equal to reactive power of the load Q (load). This happens due to the fact that the DG inverter will seek a frequency where the current–voltage phase angle of the load equals that of the DG inverter. This is because:

$$P\left(load\right)=\left(V^2 \left(PCC\right)\right)/R,$$

(3)

$$\mathrm{Q}\left(\mathrm{load}\right)=\mathrm{V}^2\left(\mathrm{PCC}\right)\left(1/\left(2\times \mathrm{pi}\times \mathrm{f}\times \mathrm{L}\right)-2\times \mathrm{pi}\times \mathrm{f}\times \mathrm{C}\right).$$

(4)

This deviation can then be effectively measured to find the islanding case. Furthermore, the detection time of the over/under voltage is found to be 0.03 and 0.05 s and that of the over/under frequency is found to be 0.02 in both cases [56,57]. The main drawback of the over/under voltage and over/under frequency anti-islanding method is its vast NDZ. This range is given by:

For active power,

$${\left(V / Vmax\right)}^{2 }- 1 \le ∆P / P \le {\left(V / Vmin\right)}^{2 },$$

(5)

For reactive power,

$$\mathrm{Qf}\left(1-{\left(\mathrm{f}/\mathrm{fmin}\right)}^2\right)\le ∆\mathrm{Q}/\mathrm{Q}\le \mathrm{Qf}\left(1-{\left(\mathrm{f}/\mathrm{fmax}\right)}^2\right).$$

(6)

These two ranges act as the limiting value of the voltage and frequency for the detection for the islanding condition in the net metering.

2.2.2. Active Method

The active method of islanding detection deals with the perturbation to the system directly. It means that an additional external variable, positive feedback or controlled change is introduced, which interact with the power system operation providing the significant change in system parameters during islanded conditions and negligible change during proper operation of the DG in grid. It involves the feedback technique or some mechanism to find islanding through parameters change [58]. This method is more accurate and effective than the passive method due to relatively small NDZs, but it has slow response toward the perturbation. In addition, the power quality and system stability may degrade due to the external variable introduced in the system [59]. However, this method requires a remarkably higher cost of implementation [49].

A. Active Frequency Drift (AFD):

In this method, a distorted output of the inverter feedback to the utility at a slightly higher or lower frequency at a point of common coupling. AFD frequency (f’) can be calculated as:

$$f^{\prime }=f\left(1/\left(1-Cf\right)\right),$$

(7)

where Cf is the chopping fraction. It is the ratio of zero conduction interval to half of the period of voltage waveform, which determines the difference between the frequency of inverter’s output current and the frequency of grid’s voltage.

The added distorted frequency is used as a factor for detection of the islanding condition; however, under normal conditions, the added distorted frequency cannot change the frequency. The weakness of this method is that it requires a decrease in the output quality.

B. Current Noticing:

The DG inverter appears as a current source to the utility:

$$i = I sin\left(w+\beta \right).$$

(8)

In this method, a continuous variation is supplied in one of these factors:

∆V = ∆P/2*(R/P)^1/2,

(9)

where P is active power and R is resistance.

At the case of islanding, there is a considerable change in the voltage. This change can be measured to detect the islanding method. This method cannot be used in the multi-inverter case. This is because, as more variation is imposed on more inverters, the total variation will be less and cannot be used for the detection islanding condition [51].

C. Sandia Voltage Shift:

This uses positive feedback to prevent islanding. A positive feedback of voltage and the DG inverter reduces its current and thus its output power. At normal operation, there is no effect when the power is reduced. However, during the absence of the utility, there is a reduction in the voltage. This reduction results in the reduced DG inverter current. These reductions can be calculated to find the islanding condition. The weakness of this method is the inverter’s decreased efficiency and a decrease in the output power quality [50].

2.2.3. Hybrid Method

With the passive method being cheap and relatively unreliable and the active method being reliable but with the trade-off in power quality, both of the discussed methods have their own set of advantages and disadvantages. The hybrid method aims to utilize the perks of both methods discussed above while suppressing their disadvantages [60,61]. The passive method is used to monitor the overall stability and occurrence of islanding, and, in the event of islanding, the active method is employed to confirm the islanding event [3]. Some techniques include:

A. Technique Based on Voltage Unbalance and High-Frequency Impedance [3,62]:

This technique continuously monitors the voltage unbalance (passive method) as follows:

$$\mathrm{Voltage}\mathrm{Unbalance}=\left(\mathrm{Positive}\mathrm{Cycle}\mathrm{Voltage}\right)/\left(\mathrm{Negative}\mathrm{Cycle}\mathrm{Voltage}\right).$$

(10)

Voltage unbalance can occur in cases like voltage spikes, switching actions and islanding. Thus, to confirm the occurrence of islanding, positive feedback (active method) is used, where the high-frequency voltage is injected into the DG control loop. Furthermore, the voltage and current at the point of common coupling will be measured to estimate the impedance at the injected frequency. In addition, if not within a threshold, the islanding event is then detected. The selected of high frequency should not be equal to any harmonics of any generating the inverter. Furthermore, to make the signal filtering easier, there must be a spectral separation between the injected frequency and the harmonics of the fundamental frequency. Any frequency meeting these two criteria can be selected. For a 60 Hz system, 350 Hz can be selected as it lies between the fifth (300 Hz) and seventh (420 Hz) harmonics.

B. Technique Based on Voltage and Reactive Power Shift [3]:

In this method, the voltage variation of the system is measured over time and a mean is calculated (passive method) and, when this value crosses a threshold, the Adaptive Reactive Power Shift (ARPS) algorithm is used as an active method. This method varies the output reactive power of the DG system to detect islanding and the results have shown this technique to be effective, while maintaining less amount of disturbances in the inverter’s grid-connected system operation.

C. Positive Feedback Frequency Shift (PFFS) and Reactive Power Variation (RPV) Method [62]:

This method continuously monitors the system using PFFS and, when the frequency deviates by 0.1 Hz, RPV is triggered to detect the islanding. Being capable of synchronizing converters with each other, this method is thus capable of detecting islanding when multiple converters are paralleled in distributed generation.

2.2.4. Machine Learning Method

This method deals with the studies of characteristics of events from the data set (main data) and uses that knowledge to find the islanding condition in test data. The decision tree approach can make a complex decision-making process into a simpler decision with a solution easier to interpret [63].

• C4.5 Decision Tree-Based Islanding Detection [64]:

Based on the C4.5 decision tree model, this method of detection first gathers sampled data obtained from the simulation result. These data are then used to construct a C4.5 classification model. Other data are then used to verify the working of the C4.5 decision tree. Each cycle of data processing is then used for islanding detection, and the result is used to improve the C4.5 decision tree itself—thus resulting in a circular loop of island detection and improvement of the detection scheme.

B. Aggressive Signal Modeling and Support Vector Machine Based Anti-Islanding Method [65]:

System parameters like voltage and current are extracted from the point of common coupling. This is then used as input to the machines and an additional advanced machine learning technique is used to determine different scenarios of islanding and non-islanding conditions. Thus, with few parameters, predictions can be made about the islanding sate of the system. With the accuracy of 98.74%, this method was able to detect the islanding condition within 50 ms of the event occurrence.

C. Support Vector Machine (SVM) Classifier Based Machine Learning Approach [66]:

SVM classifiers are fed multiple features extracted from the system parameters such as voltage. The classifiers then determine the islanding condition of the system. Linear, polynomial and Gaussian Radial Basis Function kernels are used to train the machine and further islanding and non-islanding data are used to fine-tune the system. The proposed method is tested with various data with all the possible combinations of active and reactive power imbalances. The resulting method yields exceptional results when compared with some passive methods of islanding detection such as ROCOF and over/under frequency.

D. Reduced NDZ Based Islanding Detection [67]:

With the use of the CART (Classification and Regression Trees) algorithm, the study reduces the non-detection zones in an islanded condition, thereby overcoming the large NDZ due to reduced power mismatch in case of multiple feeders network. CART is a decision tree building algorithm that uses if-then logical conditions to predict accurate conditions of cases. Without prior knowledge of relationships between multiple variables and theories governing them, this method can be effectively used for data mining and data’s implementation to predict various cases [68]. With the reduction of 54% of total NDZ area, higher dependability and improved overall performance are achieved, though extensive field experiments are required before any practical device can be realized.

2.2.5. Remote Methods

The remote method of islanding detection is the communication-based method between the utility and DG. The islanding is known through the status of the utility circuit breaker, and a trip signal is sent to the DG unit [69]. With no NDZ regions, no effect on the power quality and fast response time, the remote method of islanding detection provides a highly dependable scheme. However, the use of extra communication channels and high-cost software like Supervisory Control and Data Acquisition (SCADA) makes this method pricey to implement. The remote islanding detection can be divided into two schemes: transfer and trip scheme and power line signaling scheme [70].

Transfer and trip scheme continuously monitor the status of all the reclosers and circuit breakers responsible for the islanded condition. Thus, a higher requirement of system interaction is essential. The SCADA system can be used for this scheme. Likewise, auto-ground is a low cost, scalable approach for anti-islanding protection of distributed generation, using a utility owned and operated system, installed just down-line of the utility protection device (substation breaker or inline recloser). In auto-ground, when the utility breaker opens, the auto-ground opens the substation side device called Sectionalizing Switch (SS) and closes the Auto-grounding Switch (AS) effectively applying a three phase to ground fault. In addition, based on online protection, all DGs connected on their anti-islanding protection will be forced to disconnect the breaker. However, in the power line signaling scheme, the power line communication scheme is used to broadcast a continuous signal from the transmission system to the distribution feeder using the transmission lines. DGs are provided with signal receivers which monitor the continuous signal and, in the case of the islanded condition, the signal disruption occurs. Some of the remote-based islanding detection methods are:

A. A Remote Islanding Detection Photovoltaic-Based Distributed Generation Systems and Wind–Solar Hybrid Power Plants [49,71]:

Monitoring multiple parameters like changes in the current through circuit breakers, frequency and voltage at PCC on three sides of the PV system, this study compares them to a pre-defined threshold. This method uses continuous communication between the control system and on-field devices like the circuit breakers. This method is independent from the local loads and inverters and thus is independent from the power mismatch. This results in zero NDZ region. Circuit Breakers are tripped in case of islanding detection. With no NDZ, the islanding detection time is found to be 1.65 m-sec for the photovoltaic-based distributed generation systems.

B. Wireless Communication-Based Islanding Detection [72]:

With the introduction of an End Monitor Module (EMM) on the load side, generation side and distribution side, current, voltage, power, status of generators, and other required parameters are measured. These measured values are then communicated with the Local Area Monitor module (LAN) using the ZigBee module. The three EMs are provided with the capability of connecting and disconnecting loads and sources. This capability of EM is used to control various devices as required at the time of islanding. The use of wireless communication ensures the modular capability of the scheme and thus does not depend on the restrictions that would have risen when using a more conventional hard-wired based communication.

A brief comparison of different islanding methods are presented in

Table 1. Comparison of islanding methods [3,49,50,51,73,74].

Technique	NDZ	Detection Time	Power Quality	Cost	Detection Reliability	Effect of Multiple DGs
Passive	Broad	Short	No effect	Low	Low	None
Active	Limited	Long	Degradation	Average	High	Synchronization issues
Hybrid	Limited	Long	Degradation	High	High	Synchronization issues
Machine Learning	Limited	Variable	No effect	High	High	None
Remote	None	Very short	No effect	High	Very high	None

below.

3. Method

There are varieties of the methods for the islanding detection of a DGS, each has its own advantages and disadvantages. In this study, two different approaches: islanding detection by the auto-grounding method and the data mining-decision tree algorithm has been presented. The detailed description of the approaches are described in Section 3.1 and Section 3.2 below.

3.1. Auto-Grounding Method

In the auto-grounding method, it is assumed that the DG’s line protection has been properly configured in order to detect all faults. In case of inverter based DG, overcurrent protection alone may not be sufficient and more advanced functions such as overcurrent with voltage restraint may be required. The auto-ground system consists of three main components: Sectionalizing Switch (SS), Auto-grounding Switch (AS), and the controller, which can be implemented using a variety of re-closer or breaker controls. The SS is required only to mirror the state of the substation breaker, and it is preferable to have the SS as a separate device for the simple reason that costs escalate when work within the substation is required. Any savings associated with integration of the SS function into the feeder breaker would be outweighed by the cost of linking it with the AS [65]. In this case, it is not required to interrupt fault currents over any device that can provide automated sectionalizer capabilities. For applications where the auto-ground is paired with an in-line re-closer, the SS is not required as its functionality can simply be integrated into the re-closer itself. Figure 2 shows the block of a system for the implementation of islanding detection via the auto-grounding method, and the following are the three logics/stages of the process:

• Both of the circuit breakers (CB1 and CB2) are in closed conditions. That is the normal condition. Thus, the voltage and current waveforms found at the utilization level are purely sinusoidal.
• CB1 is initially closed and is opened with transition time ‘t’, but CB2 is closed. After switching off the CB1, the whole part of the distribution system is isolated from the remaining part of the power system, which is called the islanding condition.
• CB1 is initially closed and is opened with transition time ‘t’. As another DG is there in the line where CB2 is connected, the opening of CB2 is the condition of anti-islanding.

3.2. Data Mining-Decision Tree Algorithm

Decision trees are a type of supervised machine learning process where the data is continuously split as per the prescribed parameters. The tree can be explained by two entities, namely decision nodes and leaves. The leaves are the decisions or the final outcomes, and the decision nodes are placed where the data is split. It can be divided into two types: Yes/No type and Continuous data type. It is one of the best independent variable selection algorithms that make a model with logistic (or linear) regressions or with neural networks, which reduce the number of variables by selecting only the relevant ones: use decision trees. The process is fast as compared to the calculation of simple correlations with the target variable, and it also takes into account the interactions between variables. Decision trees are non-parametric means and no specific data distribution is necessary, and easily handle continuous and categorical variables.

The main importance of decision trees is that they identify subgroups, in which each terminal or intermediate leaf can be seen as a subgroup/segment of your population. It delivers high-quality models, which are able to squeeze pretty much all information out of the data, especially if you use them in ensembles. In this study, this concept is used to detect the islanding state of a DGS.

The data from the 13-bus system is extracted into the tree model by using the Classification and Regression Tree Algorithm (CART). The CART algorithm uses the predictive modeling for designing the decision tree. CART was first developed by Breiman et al. in the 1980s [75], and was first introduced in the Power System by Wehenkel [76,77]. The CART model of the DGS is obtained by recursively partitioning the data space and fitting a simple prediction model within each partition, due to which partitioning can be represented graphically as a decision tree [78,79]. CART uses the empirical impurity reduction known as Gini Index for splitting the node and selects the split that maximizes the decrease in impurity and generates the sequence of sub-trees by growing a large tree and pruning back until only one root node is left [80]. The Gini index is given by:

Gini (t) = 1 − Σ [p(j|t)]²,

(11)

where p(j|t) is the relative frequency of class j at node t.

The decision tree model generated using the CART algorithm is tested in Python software using tested and validated test data obtained from the auto-grounding method.

4. Result and Discussion

The match between the source and loads lead to the anti-islanding methods being unable to detect the islanding conditions. The closer the power match between the loads and the sources, the lesser the changes of the various parameters (like voltage change, frequency change, etc.) occur during the islanding event. This means that the anti-islanding methods which depend on monitoring these parameters will tend not to detect the islanding event. To overcome this, monitoring more than one parameter is required.

For the developed method, the power mismatch between the source and the loads are 38.41% to −24.39% as shown in Figure 3. That is, the case of any power mismatch between these limits will not be detected by the proposed methods. This NDZ has to be determined by conducting the load flow analysis of a sample of 13-bus system as shown in Figure 4.

4.1. Result of Islanding Detection by the Auto-Grounding Method

Auto-ground, anti-islanding protection is based on applying a three-phase, short-circuit to the islanded distribution system just prior to reclosing or re-energization. It is seen as a compromise in terms of cost and performance between the transfer-trip and local passive measurements. The auto ground system is installed just down-line of the utility protection device (substation breaker or inline recloser). In this configuration, following opening of the utility breaker, the auto-ground opens the substation side device, denoted SS and closes the AS effectively applying a three phase to ground fault. All DGs that have not already disconnected based on their anti-islanding protection will be forced to disconnect based on online protection.

For the detection of islanding events, four passive methods are used; over/under frequency, over/under voltage, rate of change of power and rate of change of frequency. The use of passive methods along with the principle of auto-grounding has been used for the detection of the islanding conditions. Based on the limits defined, the state of the system was monitored for 42 distinct cases.

Among the islanding (1) and non-islanding (0) states with the 42 sequences of events, half were in islanding state, while the other half of the events were not in islanding state as shown in

Table 2. Forty-two individual events for the islanding detection using the auto-grounding method.

No.	Frequency	Voltage	df/dt	dP/dt	Islanding States
	(Hz)	(pu)	(Hz/sec)	(MW/sec)
1	49.977	0.911	0.748	44.039	1
2	50.279	0.989	7.309	508.538	1
3	49.986	0.935	0	1.981	0
4	49.979	1.05	0.175	16.706	1
5	49.986	0.969	0	1.15	0
6	49.986	0.953	0	1.127	0
7	50.043	0.95	2.364	43.767	1
8	50.216	1.2	1.967	1974.419	1
9	50.108	0.987	0.738	1.98	0
10	50.045	1.016	2.294	18.755	1
11	50.108	0.972	0.738	1.127	0
12	49.986	0.954	0.738	0.551	0
13	49.803	0.984	4.868	49.238	1
14	49.77	1.174	5.684	35.284	1
15	49.997	0.862	0.007	2.391	0
16	49.928	1.106	1.72	12.699	1
17	49.997	0.914	0.007	2.46	0
18	50.022	0.94	0.619	10.169	0
19	49.977	0.902	0.233	35.521	1
20	49.904	1.006	2.052	11.327	1
21	49.986	0.94	0	0.328	0
22	49.982	1.038	0.106	6.963	1
23	49.986	0.977	0	0.282	0
24	49.986	0.96	0	0.639	0
25	49.952	0.916	0.857	48.117	1
26	49.953	1.025	0.837	21.89	1
27	49.986	0.931	0	2.799	0
28	49.957	1.059	0.729	15.551	1
29	49.986	0.964	0	1.653	0
30	49.986	0.948	0	1.111	0
31	49.977	0.908	0.233	42.504	1
32	50.002	1.017	0.39	22.384	1
33	49.986	0.939	0	2.203	0
34	50.167	1.002	4.51	16.962	1
35	49.986	0.972	0	0.29	0
36	49.986	0.957	0	0.596	0
37	49.952	0.912	0.857	44.299	1
38	49.953	1.02	0.837	17.891	1
39	49.986	0.932	0	2.07	0
40	49.953	1.023	0.837	8.074	1
41	49.986	0.967	0	1.598	0
42	49.986	0.951	0	1.165	0

. The developed auto-grounding method was implemented to detect these conditions. Furthermore, 20 of the total 21 islanded conditions (95.2%) and 19 of the total 21 non-islanded conditions (90.4%) were successfully detected, and listed in

Table 3. Accuracy of the model with the auto grounding method.

Class	Cases	Correctly Classified
Islanded Condition	21	20 (95.2%)
Non-Islanded Condition	21	19 (90.4%)
Total	42	39 (92.8%)

.

For instance, the event number 3 (with frequency 49.986 Hz, voltage of 0.935 pu, rate of change of frequency of 0 Hz/sec and rate of change of power of 1.981 MW/sec) was detected to be in a non-islanding condition, while event number 7 (with frequency 50.043 Hz, voltage of 0.95 pu, rate of change of frequency of 2.364 Hz/sec and rate of change of power of 43.767 MW/sec) was detected to be in an islanding condition.

Figure 5 shows the current and voltage waveforms in the event of islanding detection by the auto grounding method. In the event of islanding in 2.5 s, it takes the auto-grounding method 1.705 s for islanding detection. With the auto-grounding scheme grounding the lines, the current increases above 500 A at around 4.194 s and thus the line protection scheme disconnects the circuit breaker, thereby creating an anti-islanding scheme.

4.2. Result of Islanding Detection by the Data Mining-Decision Tree Algorithm

Similarly, on the other hand, the decision tree is generated using the CART algorithm in the scikit-learn module of the Python programming language [81]. From the total data set, 70% of the data are used as training data and the remaining 30% are used as testing data. The testing data set are generated randomly for each case of islanding state and non-islanding state. The topmost part of the tree is also known as the decision node, which consists of the condition, can be either true or false, following the path of true or false, and it finally ends at a leaf node, while, for a leaf node, the ‘value’ consists of two pieces of numeric data. The value on the left-hand side represents the value of the non-islanding condition and on the right-hand side represents the value of the islanding condition.

The total training samples are 88, out of which 44 samples are for the islanding condition and another 44 are for the non-islanding condition. The node is split until the minimum gini for the samples are obtained and all the samples are correctly classified. For instance, the data sample where the value of active power is calculated as 1, reactive power as 0.64, rate of change of power (Dp) as −7.89, voltage as 0.9857 pu and power angle as 25.25 degrees is taken to explain the workings of the developed algorithm. In accordance with the developed tree, first the algorithm checks the power angle, which is greater than 21.75; then, it proceeds to check power (P) which in this case is less than 34.54. The condition is true and it proceeds towards the left side to check if the power angle is greater than 22.316. This is false, which directs the algorithm to check Dp/P, which in this case is less than 140.111. Furthermore, it checks if the power angle is less than or equal to 22.99. In the same manner, the process continues and finally ends at the leaf node at the bottom left corner, thus determining the condition to be the ‘islanding condition’. This path is denoted by the darker shaded decision boxes in Figure 6. The gini index in the box is used to determine the uncertainty in the result. Here, the gini index is zero, meaning that the result obtained is 100% certain. The actual tree diagram of the conducted analysis is shown in Figure 5.

The entire data set was classified into islanding mode and non-islanding mode. The model successfully classifies all the training set data to the correct classification as shown in

Table 4. Accuracy of the model.

Classes	Training Set		Testing Set
	Total Cases	Correctly Classified	Total Cases	Correctly Classified
Islanding	44	44(100%)	18	17 (94.4%)
Non-Islanding	44	44(100%)	19	18(94.7%)
Total	88	88(100%)	37	35 (94.5%)

. Thus, the accuracy of the model with the training set data is 100%. Likewise, for the testing set data, the model classified 35 out of 37 events to the correct classification. Thus, the accuracy of the model for testing set is 94.5%. In one of the cases, the occurrence of islanding in 2.5 s led to the detection of islanding within 0.162309 s. Simulation of both the methods were conducted using the computer having Intel^® Core ™ i7-7500U @ 2.70 GHz (4 CPU), ~2.9 GHz processor specifications and 8192 MB of RAM.

4.3. Comparison between Auto-Grounding Method and the Data Mining-Decision Tree Algorithm Method

Based on the two simulations, the two islanding detection methods are compared in accordance with their respective accuracies in Figure 7. While the auto-grounding method is more accurate in detecting islanding conditions with 95.2% accuracy compared to 94.4% accuracy of the data mining-decision tree algorithm method, the data mining-decision tree algorithm method is vastly more accurate in predicting the non-islanding conditions with the accuracy of 94.7% compared to the accuracy of auto-grounding method being only 90.4%. Taking both islanding and non-islanding conditions, the auto-grounding method is only 92.8% accurate while the data mining-decision tree algorithm method is 94.5% accurate. Auto-grounding takes a long time to detect the islanding condition (1.705 s), while the data mining-decision tree method is quicker to detect the islanding condition (0.162309 s).

5. Conclusions

The accurate and precise detection of the islanding condition plays a very significant part in DGS. This paper highlights a relatively new approach of using the auto-grounding method and data mining decision tree technique to correctly detect the islanding condition of DG. The low-cost auto-grounding method was designed and simulated, and it was observed that the accurate anti-islanding can be done within 1.706 s. Moreover, the use of the data mining decision tree algorithm based on a machine learning approach successfully identified the anti-islanding condition with high accuracy and reliability within 0.162309 s. The comparison between auto-islanding and the machine learning technique shows that using the data mining decision tree technique presents the fastest operation of the anti-islanding condition. Furthermore, the data mining decision tree approach was also found to be more reliable and accurate than the auto-grounding method.

An auto-grounding method, which is a type of remote method for islanding detection, and a data mining-decision tree method, which is a type of machine learning methods for islanding detection, were developed, simulated and evaluated. Auto-grounding methods used four passive methods for islanding detection: over/under voltage, over/under frequency, rate of change of power and rate of change of frequency. Forty-two individual events were used to simulate the islanding detection using the auto-grounding method, out of which 21 events were in islanding state and 21 events were in non-islanding conditions. Similarly, 37 events were used in the data mining-decision tree method, where 18 events were in islanding states and 19 events were in non-islanding states.

The proposed islanding detection methods were simulated in the computer and the following conclusions were derived:

• The data mining-decision tree method has a higher overall accuracy (94.5%) when compared to the auto-grounding method (92.8%). Thus, comparing between the two, the data mining-decision tree method is the more reliable method of islanding detection.
• In comparison between the two discussed methods, the data mining-decision tree method presents the fastest and more accurate operation of anti-islanding condition.

Auto-grounding being the older and the data mining-decision tree method being the comparatively newer method of islanding detection, these simulations highlight increased accuracy and speed of the newer islanding detection method when compared to the older one. The power grid system’s complexity is growing on a yearly basis. The introduction of DGs and increased power capacity of the power grid has led to the requirement of a faster protection system. Two seconds is not a fast-enough islanding detection speed in this scenario. As such, micro-grids and smart-grids are relying more on newer and faster islanding detection methods.

In successful simulations of the data mining-decision tree method, the research has shown, with positive results, a newer form of islanding detection method with short detection speed and high accuracy. This method promises to satisfy the meticulous islanding detection criteria of the future grids comprised of micro-grids and smart grids. To further the findings of this paper, the following explorations can be conducted:

• The detections of islanding conditions were done in 1.705 s and 0.162 s by auto-grounding and data mining-decision tree methods. The speed of islanding condition detections can be improved upon by simulating these two methods in a computer with better specifications.
• A larger amount of training data can be fed to a data mining-decision tree method to further increase the accuracy of this method.
• Expanding on the presented method by making it compatible with the grid code. This may include and not be limited to introducing fault ride through requirements.