Deep learning for islanding detection of grid-connected photovoltaic systems
Over the past decade, the integration of renewable-based distribution energy resources within the smart distribution system has been steadily growing. Despite the numerous advantages of integrating these renewable energy resources in reducing the greenhouse gas emissions and releasing the transmission line capacity, there are many operational challenges involving the protection when considering the unintentional islanding of these resources. Typically, when faults occur, an island or multiple islands may be formed throughout the electrical power distribution system. When the island includes one or more distributed energy resource, there is usually a serious potential hazard for the working personnel and hence the IEEE Standard 1547 recommends the distributed generation to cease to supply within 2 seconds of the formation of the island. Thus, there is a need for detecting the island and disconnecting the distributed generation within two seconds to fully comply with the IEEE Standard. In the literature, several methods for islanding detection have been proposed, which can be classified as communication-based, active or passive methods. The communication-based methods are expensive to implement while the active methods typically inject harmonic distortion into the distribution system to detect the island, which will lead to power quality degradation. On the other hand, the passive methods are preferred over other detection methods because they are inexpensive to implement and they do not affect the system power quality. The passive methods rely on identifying the islanding features in the local measurement of the voltage and current signals, which needs to be fed to the machine learning algorithms that have been proposed in the literature. However, identifying such features a priori is a very complex task in particular when considering inverter-based renewable energy resources. The work in this thesis uses deep learning to discover the features of the islanding as part of the classification process rather than identifying the features a priori as in the previous work and hence enhancing the capability of the passive methods. The results of testing the proposed passive islanding approach show an outstanding performance of the proposed method in islanding detection in grid-connected photovoltaic system. The proposed method utilized 31 features generated from 46 different cases to classify and detect the islanding. The proposed method succeeded in detecting islanding with a high accuracy of 98.1% and a rapid detection time of 0.28 second.