Islanding Detection for Inverter-Based Distributed Generation using Unsupervised Learning /

Abstract

Islanding detection with the rising grid supporting inverter-based distributed generation is becoming more critical protection due to its high droop gains and overall decreased system inertia leading to rapid changes in the electrical parameters. Traditional methods for islanding detection in this regard are susceptible to significant problems such as nondetection zone, false-positive detection, and inefficient mode of validation. Therefore, in order to attenuate these problems, this paper proposes a hybrid islanding detection technique based on unsupervised anomaly detection using autoencoders. This technique uses the rate of change of frequency, and phase angles of the voltage and current as primary and secondary detection parameters which demonstrates improved performance, reliability, and robustness due to its shared advantage of both, active frequency drift and autoencoder. Furthermore, a dialectic model of offline and online validation schemes is also proposed to ensure the reliability of detection. Extensive simulations and validations have been carried out on multiple networks in order to generate data-sets that were used to train, test, and validate the technique and compute its statistical significance thereby confirming its effectiveness. The optimal islanding detection time for the base cases was recorded as 20 milliseconds with an F1-score of 0.991, dependability index of 0.998, and security index of 0.995, with zero non-detection zones, which complies with IEEE standard 1547’s requirement of detection within two seconds after islanding.