Islanding detection for microgrids using machine learning techniques

The recent emphasis on using renewable energy instead of the traditional power generation techniques, which are badly affecting the environment, has introduced the use of Distributed Generation (DG) power grids. In these grids, smaller energy sources are distributed over the entire grid, instead of using larger centralized power units as usual. This type of grids imposes new challenges in operating and protecting the power generation units, as well as the quality and availability of the power flowing through the grid. One of the most important challenges faced by using the DG power grids is islanding detection, where a specific part of the DG grid, which is known as a microgrid becomes energized by the power sources in that part, but not connected to the main power grid. Such situation must be detected in order to disconnect DG power sources from the distribution grid, to protect the DG distribution systems from the negative influence of such situation on the management, protection and operation of these distribution systems. The IEEE standard for distributed generation power grids recommends detecting islanding events in DG grids within an interval of two seconds. There are many methods proposed to detect an islanding event. These methods can be divided into three main categories that are active, passive and communication-based methods. The active methods are based on inserting some distortion to the output form of one of the DG power sources, usually inverters. The effect of this distortion is then measured on the terminals of that device. In case an islanding event occurs, the effect of that distortions becomes more noticeable, as the power grid becomes smaller. Active methods have a good capability of quickly detecting an islanding event, but the equipment used in these methods are expensive and they cause degradation to the quality of the power flowing in the grid. Communication-based methods use interchanges information between different parts of the power grid in order to detect an islanding state. This information includes values that are remotely measured for certain nodes in the DG grid, such as the breakers status, voltages, power and frequency. These methods are also capable of detecting an islanding event quickly and accurately, but they require intensive communication systems, which are relatively expensive systems, and the reliability of the protection system depends on the availability of the communication system, where a failure in the communication system leads to a failure in the islanding detection system. Thus, although these methods do not degrade the quality of the power flowing in the DG grid, it requires an expensive infrastructure, and they are highly dependent on the availability of that infrastructure. Passive islanding detection methods, on the other hand, are installed locally, require no communication infrastructure, and do not degrade the quality of the power flowing in the DG grid. These methods rely on monitoring specific values in the DG grid in order to detect an islanding state. The existing methods have limited detection capability, compared to the methods of other categories. But the low cost of building such method and the independence from any other system make methods from this category promising. The only drawback in the existing methods of this category is the wider, compared to the other two categories, Non-Detection Zone (NDZ), which is caused by the less capability of detecting some islanding events. In this study, the performance of a new passive technique, which relies on using convolutional neural networks, is tested. Convolutional Neural Networks (CNN) simulates the way living creatures see the surrounding environment. They are based on detecting the existence and location of features that are detected during the training of the neural network. Three matrices are created based on the last predefined number of instantaneous value for the voltage of each phase. These matrices are fed to the neural network in order to predict whether the DG grid is in normal or islanding state. The use of instantaneous values enables the CNN of faster detecting for the islanding events, unlike other techniques that use values such as frequency and RMS values, which represent a summary of the instantaneous values of the measured variable for a specific period of time. Sudden changes require a longer time to affect such values, which leads to longer detection time. On the other hand, some methods rely on summarized values, such as the rate of change, which require a longer time to reflect the change in the measured value. A test model is built using MATLAB & Simulink, in order to simulate different, islanding and non-islanding, events. Then, different classifiers are built based on the data retrieved from the simulated events, as these classifiers could be used to detect such events. Finally, the convolutional neural network is built using python programming language and Tensorflow library for neural networks implementation. All experiments are executed using an Intel® Core™ i7-4500U CPU @ 1.8GHz and a memory of 8.00 GB capacity with 2.00 GB of external GPU memory, running Windows 10 operating system. The use of CNN in islanding detection shows better performance than the other classifiers, where the accuracy achieved by the CNN is 99.78 and a detection time of 11.7 milliseconds, while the decision tree classifier achieved 97.02% accuracy and 61.64 milliseconds detection time, random forests achieved 99.08% accuracy and 19.01 milliseconds detection time. The k-NN classifier achieved a very high accuracy of 99.97%, but the relatively high execution time of this classifier produced an overall detection time of 50.53 milliseconds. The Bayes Net classifier achieved 99.34 accuracy and 14.24 detection time. These tests are conducted by excluding four events, two islanding and two non-islanding, from the training dataset and use them as test dataset to evaluate the performance of these classifiers. The overall time consumed by the CNN to detect an islanding event is 11.7 milliseconds, where the CNN required 4.5 milliseconds to sense the islanding event, and 7.2 milliseconds for the computations. This computation time can be reduced by using computers with better hardware specifications, or by using some special accelerators in order to achieve faster islanding detection.