Intelligent method for islanding detection of photovoltaic distributed generators based on voltage waveform recognition

The anti-islanding protection of distributed generators (DGs) is typically performed by conventional schemes that monitor the magnitude and frequency of voltage signals. However, one of the main issues in setting up such schemes is to identify and differentiate magnitude and frequency variations caused by islanding events from other possible disturbances, such as voltage sag or swell, flickers, harmonics, among others. In this context, an algorithm based on Artificial Neural Networks (ANNs) can be used to identify patterns in voltage waveforms, allowing it to differentiate islanding events from other types of disturbances. A good performance was observed by using the intelligent proposed method that presents a high hit rate against different scenarios and operating conditions. Nevertheless, the ANN training process is not simple, since it involves the definition of the ANN architecture, the data window length, the sampling rate and the selection of a representative training set for the analyzed power grid. Therefore, a systematic approach is presented in this paper to specify and train ANNs for islanding detection of photovoltaic DGs. The results show that, among more than 200 different tested ANN architectures, the best one with two hidden layers achieves a precision of 99.5% and an F0.5 score of 99.2%. A real scenario was modeled for developing and testing the proposed method, while the simulations were performed in Matlab. Considering nearly 600 simulations of islanding events, a high hit rate was observed, depending on the detection time, with 67% being detected within 0.07 s. Additionally, the proposed method was consistently able to differentiate islanding events from voltage and frequency variations, even with variations of up to 18% and 10% of their nominal values, respectively. All relevant aspects related to the proposed method are discussed in this paper.