Monitoring of Synchronization Failure for Power Electronics Converters

The recent developments in the grid integration of power electronics converters have raised some serious concerns related to their synchronization. These concerns when unattended resulted in unintentional islanding of the distributed generation (DG) units and triggered the cyclic behavior of system disconnection. Hence, to overcome these drawbacks, timely detection of these synchronization failure with high accuracy is necessary. In this paper, the applicability of an advanced recurrent neural network that allows persistent information exchange is analyzed for distinguishing between the normal operation and synchronization failure in a converter network. To achieve this a long-short term memory (LSTM) autoencoder is designed as a classification approach and employed with the power converters operating in a network. The LSTM offers advantages with automated feature extraction and ranking which are the major aspects for improving the time detection of disturbances in a system with high accuracy. To develop this approach, a three-phase grid feeder integrating a three-phase rectifier with a three-phase inverter and a single-phase inverter is designed. Measurements and normal operation and synchronization failure are analyzed to train the algorithm. The trained algorithm is identified to achieve 99.49% training accuracy and 100 % testing accuracy with a detection time of 0.2 milliseconds.