The growing use of distributed generation instead of the traditional localized larger generation units is imposed by the employment of renewable energy generators, which are usually of smaller sizes distri buted over the power grid. The members of the European Union, for example, are targeting to energize 20% of the required power using renewable power sources [1]. For optimal operation, power grids that use distributed generation are divided into smaller grids, known as microgrids, that are interconnected to each other, where the power transferred from one section to another are kept at minimum possible value, where zero transferred power among microgrids is considered the optimal value. However, these microgrids are still required to be connected to the main grid to maximize the stability and the quality of the power in the grid [2]. A microgrid is described to be in an is landing state when the power generation units are energizing the loads in that microgrid, while the entire microgrid is disconnected from the main power grid. As recommended by the IEEE 1547 standard [3], the power sources in an is landed microgrid must be disconnected as fast as possible within a maximum time window of two seconds. The aim of disconnecting these power sources is to protect them from any sudden load change that may be severe compared to their limited power rating and maintaining the quality of the power provided to the customers, which is controlled by the main grid.

Thus, many techniques are proposed to detect is landing events in order to disconnect the power sources in a microgrid.
Passive islanding detection is one of the most inexpensive techniques that does not degrade the quality of the power in the grid. However, the existing techniques suffer from low accuracy in distinguishing is landing from non-is landing events, so that, the system may disconnect the power sources upon false alarms that may occur when non-is landing events occur in the grid, or may keep energizing the grid even if an islanding event occurs [4,5]. In this study, a new passive is landing detection system is proposed, which makes use of artificial neural networks. The proposed system is implemented to detect is landing events from the instantaneous value of the voltages in the grid, so that, more accurate and rapid decisions can be made. According to the existence of multiple types of neural networks that can achieve such a task, these types are evaluated in order to select the most appropriate neural network for this purpose.

The remaining sections of this article are organized as follows: Section 2 reviews literature related to is landing detection methods and artificial neural networks; Section 3 illustrates the proposed is landing detection technique. Section 4 describes the experiments conducted to evaluate the proposed is landing detection system using the different types of artificial neural networks. Section 5 concludes this article and suggests improvements that are considered in future work.

Islanding detection techniques can be categorized into three main categories, which are the active, passive and communications-based techniques. Active techniques rely on injecting noise to the waveform generated by the power source, mainly inverters, then measure the effect of that noise on the actual waveform in the power grid. The smaller the grid that the noise is injected to, the higher the effect of the noise is. Thus, active islanding detection techniques have a narrower non-detection zone. However, as the share of power sources that use active islanding detection is increased in the power grid, the effect of the noise they inject increases, which reduces the quality of the power provided to the customers [6,7].

Passive islanding detection techniques monitor one or more variables of the power provided on the grid, such as frequency or power angle, to detect islanding events. These techniques are of lowest cost, compared to the other islanding detection types, and do not degrade the quality of the power provided to the customers. However, the values used by these techniques to detect islanding events are summaries of values calculated over a certain interval, such as the use of root-mean-square for voltage and current and phase loop lookups for frequency. As these values are calculated based on multiple instantaneous values, a change in these values does not appear instantly, as the older values still dominate the computed values. Moreover, detecting a change in the microgrid when the islanding occurs when no energy is imported from or exported to the main grid is difficult, as no sudden changes are going to happen to any of the monitored values [8,9]. Communications-based techniques interchange information about the status of different circuit breakers in the power grid, so that, any combination of failures that may cause islanding events are detected instantly. These techniques require an expensive communications infrastructure in order to exchange such information. Thus, these techniques rely completely on the communications infrastructure, which reduces the reliability of such techniques, as the failure of the communications system causes a loss of the islanding detection system. Moreover, these techniques are the most expensive among the other islanding detection types [10].

The active islanding detection techniques proposed in [11–13] use Sandia-Frequency Shift to add noise to the power generated by the power sources in the microgrid. These methods are normally implemented in power sources that use inverters, as the addition of noise to the power generated by a rotating power source if more difficult. When the microgrid is connected to the main grid, the injected noise is absorbed by the rotating generators, as these generators are of higher capacities. Thus, by measuring the effect of the Sandia-Frequency Shift at the terminals of the inverter, connected to the microgrid, it is possible to estimate the size of the grid that the inverter is connected to, so that, when the effect rises over a certain threshold value, the microgrid is considered to be working in islanding state and power sources are shut down [14]. However, the Sandia-Frequency Shift still have a degrading effect on the supplied power, even if it is relatively small, but such effect is amplified when the number of power sources that use such method is increased in the grid. In the existing passive islanding detection techniques, Rate of Change of Frequency (ROCOF) is one of the widely monitored values. In normal operation conditions, when a change in the loading of the grid occurs, all power generators in the grid participate in the compensation of the power required to cover that load, so that, the frequency of the provided power does not drop dramatically. Moreover, a reduction in the loading of the grid is also compensated by all the power generators, by reducing the provided power, to maintain a stable frequency. Thus, higher ROCOF indicates that the microgrid is operating in islanding state, so that, all power sources should be shut down [15]. In addition to the time required by the phase-loop lookup to measure the frequency and the time required to calculate the ROCOF, such systems may misfire in cases of rapid sudden load changes or loss of larger generation units, which eventually degrades the quality of the power provided to the customers.

The aim of the proposed method is to analyze the instantaneous voltage values to detect any distortion that indicates the occurrence of an islanding event. As the values of these voltages follow a sinusoidal distribution, the detection of such distortions requires the use of time-series analysis techniques. With the ability of Artificial Neural Networks (ANN) to outperform many of the state-of-the-art methods in different applications, the proposed method relies on these networks to detect islanding event. However, to maintain the ability to sense changes in the values with respect to time, the proposed method uses Recurrent Neural Networks (RNN). The main three types of RNNs are evaluated in this study, which are the simple RNN, Long- Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU).

3.1. The Simple RNN

In this type of RNNs, the output of the neuron at a certain time instance is weights and feedback to the neuron, as input, in next time instance, as shown in Figure 1. Accordingly, the features detected by this neuron is constructed from a set of sequential values. However, as the values of the output are simply multiplied by a weight value, the older values that appear in the time-series have less influence on the value being calculated, while recent values have significantly higher influence. Despite this limitation, the sinusoidal nature of the input values and the simplicity of this type of RNN neurons has encouraged the evaluation of this type of RNNs, as the recent values in a sine wave can be used to recognize any upcoming distortions and the simplicity of the computations allows faster detection of such distortions.

Figure 1. Computations in a simple RNN cell.

3.2. The Long- Short-Term Memory (LSTM)

In order to allow RNNs to adjust the influence of values based on their importance in the feature being detected, rather than their relative position to the current time instance, LSTM uses a set of gates to control the flow of these values in the computations, as shown in Figure 2. These gates allow the neuron to adjust the influence of the value, regardless of its position, so that, older values can have higher influence on the computations in the current time instance. However, the complexity of this type of RNNs is significantly higher than the complexity of the simple RNN, according to the significantly higher number of mathematical operations required by the gates.

Figure 2. Computations in a LSTM cell. 3.3. The Gated Recurrent Unit (GRU)

To reduce the complexity of the RNN, while maintaining the ability to rely on the importance of the value in the computations, rather than its relative positioning, the GRU uses the hierarchy shown in Figure 3. The structure of this unit is significantly less complex than the LSTM, whereas the existence of the remember and forget gates maintain the ability to control the flow of computations during the computations of a certain input at a certain time instance. Hence, this type of RNNs has a relatively lower complexity than the LSTM but maintains the ability of controlling the influence of the data at each time instance, similar to the LSTM. Moreover, compared to the simple RNN, the cell in GRU provides better control with slightly more complexity in the computations, compared to the LSTM with respect to the RNN.

Figure 3. Computations in a GRU cell.

3.4. DataCollection

In order to generate a data set to train and evaluate the performance of the selected machine learning technique, the IEEE 13-node test feeder model [16], shown in Figure 4, is implemented in Matlab, where instantaneous voltage values are collected for the three phases of the power grid. The events shown in [4] are included in the simulations, where some of these events cause is landing state, while others do not. A label is provided per each instantaneous value that indicates the state of the microgrid at that instance, which is used for the training and evaluation of the classifiers.

Figure 4. IEEE 13-node test feeder model [4]. 3.5. Structure of the Neural Network

The last 100 instantaneous voltage values are collected from each phase, which creates an array of the shape (100, 3) as the input of the neural network. Hence, the input layer of the neural network has the same shape and is followed by an RNN layer with 32 neurons. To avoid overfitting the neural network to training data, dropout layers are used after each hidden recurrent layer, except the last hidden layer, which is not followed by a hidden layer and fully connected to the output layer. The dropout rate at these layers is set to 20%, i.e., 20% of the neurons in each layer is dropped during training. The first RNN layer is followed by another RNN layer, with 16 neurons. Then, two additional hidden fully-connected, i.e., dense, layers are placed in the neural network with 16 neurons each and Rectified Linear Unit (ReLU) activation function, before placing the output layer, which has a single neuron with the Sigmoid activation function, as the neural network is implemented for binary classification, islanding or non-islanding. Figure 5 shows the structure of the Simple RNN, whereas the other networks have the same structure but use different type of recurrent neurons.

Figure 5. Structure of the implemented Simple RNN neural network.

In order to evaluate the ability of the different types of the recurrent neural network to detect islanding events using the instantaneous voltage values, the IEEE 13 node distribution feeder is implemented in Matlab’s Simulink to simulate the events described in Table 1. Each event is simulated under 24 load distributions and solar irradiation levels of the Photovoltaics (PV) array [17]. The classifiers are implemented using Python programming language [18] using Tensorflow [19] and Keras [20] libraries to implement and train the neural networks. All evaluations are conducted using 10-fold cross-validation to measure the accuracy, Dependability Index (DI) and Security Index (SI).

Table 1. Simulated events and their islanding status.

Status Islanding Islanding Islanding Islanding Islanding Non-islanding Non-islanding Islanding Non-islanding Non-islanding Non-islanding Non-islanding Non-islanding Non-islanding Non-islanding Non-islanding Non-islanding Non-islanding

The results of the conducted experiments, summarized in Table 2, show that the GRU has been able to achieve the best performance, in terms of quality of predictions. As summarized in Figure 6, there has been significant improvement in the accuracy of the predictions, when the GRU is used for the classification neural network, especially regarding the SI, which is one of the most important measures as the lower the value of the SI is, the higher risk to human lives, especially workers that are maintaining the disturbance that occurred and led to the islanding event.

Figure 6. Quality of the predictions collected from the neural networks.

In addition to the better predictions collected by the GRU, and despite the higher execution time, as shown in Figure 7, that is consumed by the GRU network to predict the state of the grid per each input, the significant improvement in the quality can justify such increment in the execution time, which is relatively small compared to the LSTM. Hence, the GRU neural network has shown a better ability to detect islanding events in a distributed generation grid, based on the instantaneous values of the voltages on the power lines.

Figure 7. Execution time of the different RNN neural networks.

The comparison shown in Table 2 compares the performance of the proposed method to the existing state- of-the-art methods, which rely on different summarized values that are collected from the grid. In addition to the better accuracy, the proposed method has been able to provide a significantly better SI, which can increase the reliability of the system, according to the importance of this factor as it directly interacts with human lives. Moreover, the proposed method has also been able to achieve a higher DI, which prevents the unnecessary disconnection of generation unit in case a non-islanding event occurs and wrongfully detected as an islanding event. Hence, better quality is provided to the consumers of the power grid, while maintaining the safety of the workers and lowering the amount of energy that is imported from the grid, according to the higher DI, which reduces the cost of such import and reduces the influence of the fossil fuel on the environment.

Table 2. Comparison of the proposed method with existing methods.

With growing use of renewable energy, new challenges are imposed by the change in the topology of the power grid, where smaller generation units are being distributed over the grid instead of larger local units.

One of the main challenges imposed by distributed generation is the detection of islanding events, where a local part of the grid becomes energized locally. According to the IEEE 1547 standard, such events must be detected within less than two seconds and all power sources must be shutdown, to protect the equipment and the individuals working on the grid. As the aim of the distributed generators is to minimize the power exchange with the grid, no significant change in the power flow. Thus, the recognition of the islanding events is a challenging task.

The performance of different recurrent neural networks is evaluated in this study, which are the Simple RNN, LSTM and GRU. Despite the low execution time required by the Simple RNN, the quality of the predictions collected from the neural network are very low, compared to the other types of RNN and existing methods. The LSTM has significantly better performance, compared to the Simple RNN but requires significantly higher execution time, compared to the RNN. This improvement in the performance is according to the ability of LSTM neurons to consider the previous values according to their importance, rather than their positioning. However, according to the lower complexity of the GRU units and their ability to control the influence of the values based on their important rather than their positioning, this type of RNNs has achieved the best performance, in terms of better quality of predictions than both the Simple RNN and LSTM, and significantly lower execution time, compared to the LSTM. The GRU has achieved an overall accuracy of 99.92%, with a 99.89% Dependability Index (DI) and 99.93% Security Index (SI).

In future work, the ability to consider other grid values, other than the voltage, is going to be investigated. However, the inclusion of additional values, increases the number of dimensions in the array that represents the input, which limits the ability of using the RNN networks. Hence, such methodology requires the use of Convolutional Neural Networks (CNN), which is another type of ANNs that has the ability to handle three- dimensional inputs.

Acknowledgments: The authors would like to appreciate patients who agree to participate and follow the study protocol patiently

Author Contributions: All authors contributed equally to the writing of this paper. All authors read and approved the final manuscript.

Conflicts of Interest: There are no conflicts of interest.

1. K. L. Anaya and M. G. Pollitt, "Integrating distributed generation: Regulation and trends in three leading countries," Energy Policy, vol. 85, pp. 475-486, 2015.

2. W. L. Theo, J. S. Lim, W. S. Ho, H. Hashim, and C. T. Lee, "Review of distributed generation (DG) system planning and optimisation techniques: Comparison of numerical and mathematical modelling methods," Renewable and Sustainable Energy Reviews, vol. 67, pp. 531-573, 2017.

3. I. Committee, "IEEE standard for interconnecting distributed resources with electric power systems," New York, NY: Institute of Electrical and Electronics Engineers, 2003.

4. R. Azim, Y. Zhu, H. A. Saleem, K. Sun, F. Li, D. Shi, et al., "A decision tree based approach for microgrid islanding detection," in 2015 IEEE Power & Energy Society Innovative Smart Grid Technologies Conference (ISGT), 2015, pp. 1- 5.

5. R. Azim, K. Sun, F. Li, Y. Zhu, H. A. Saleem, D. Shi, et al., "A comparative analysis of intelligent classifiers for passive islanding detection in microgrids," in 2015 IEEE Eindhoven PowerTech, 2015, pp. 1-6.

6. P. Du, Z. Ye, E. E. Aponte, J. K. Nelson, and L. Fan, "Positive-feedback-based active anti-islanding schemes for inverter-based distributed generators: basic principle, design guideline and performance analysis," IEEE transactions on power electronics, vol. 25, pp. 2941-2948, 2010.

7. J. Zhang, D. Xu, G. Shen, Y. Zhu, N. He, and J. Ma, "An improved islanding detection method for a grid-connected inverter with intermittent bilateral reactive power variation," IEEE transactions on power electronics, vol. 28, pp. 268-278, 2013.

8. A. Samui and S. Samantaray, "Assessment of ROCPAD relay for islanding detection in distributed generation," IEEE Transactions on Smart Grid, vol. 2, pp. 391-398, 2011.

9. S. Salman, D. J. King, and G. Weller, "New loss of mains detection algorithm for embedded generation using rate of change of voltage and changes in power factors," 2001.

10.  Q. Cui, K. El-Arroudi, and G. Joos, "Islanding Detection of Hybrid Distributed Generation under Reduced Non- detection Zone," IEEE Transactions on Smart Grid, 2017.

11.  H. Zeineldin and S. Kennedy, "Sandia frequency-shift parameter selection to eliminate nondetection zones," IEEE Transactions on Power Delivery, vol. 24, pp. 486-487, 2009.

12. H. Zeineldin and S. Conti, "Sandia frequency shift parameter selection for multi-inverter systems to eliminate non- detection zone," IET Renewable Power Generation, vol. 5, pp. 175-183, 2011.

13.  A. Hatata, E.-H. Abd-Raboh, and B. E. Sedhom, "Proposed Sandia frequency shift for anti-islanding detection method based on artificial immune system," Alexandria Engineering Journal, 2017.

14.  H. Berger, "Über das elektrenkephalogramm des menschen," Archiv für psychiatrie und nervenkrankheiten, vol. 87, pp. 527-570, 1929.

15.  L. Wu, Y. Liu, D. Zhou, J. Guo, and Y. Liu, "Observation of inertial frequency response of main power grids worldwide using FNET/GridEye," in Power and Energy Society General Meeting (PESGM), 2016, 2016, pp. 1-5.

16.  W. H. Kersting, "Radial distribution test feeders," in Power Engineering Society Winter Meeting, 2001. IEEE, 2001, pp. 908-912.

17.  A. Mellit and A. M. Pavan, "A 24-h forecast of solar irradiance using artificial neural network: application for performance prediction of a grid-connected PV plant at Trieste, Italy," Solar Energy, vol. 84, pp. 807-821, 2010.

18.  M. F. Sanner, "Python: a programming language for software integration and development," J Mol Graph Model,vol. 17, pp. 57-61, 1999.

19.   M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, et al., "Tensorflow: A system for large-scale machine learning," in 12th USENIX symposium on operating systems design and implementation (OSDI 16), 2016, pp.265-283.

20.  F. Chollet, "Keras: The python deep learning library," Astrophysics Source Code Library, 2018.