This paper proposes an AI-based approach for islanding detection in active distribution networks. A review of existing AI-based studies reveals several gaps

including model complexity and stability concerns

limited accuracy in noisy conditions

and limited applicability to systems with different types of resources. To address these challenges

this paper proposes a novel approach that adapts the WaveNet generator into a classifier

enhanced with a denoising UNet model

to improve performance in varying signal-to-noise ratio (SNR) conditions. In designing this model

we deviate from state-of-the-art approaches that primarily rely on long short-term memory (LSTM) architectures by employing 1D convolutional layers. This enables the model to focus on spatial analysis of the input signal

making it particularly well-suited for processing long input sequences. Additionally

residual connections are incorporated to mitigate overfitting and significantly enhance the model's generalizability. To verify the effectiveness of the proposed scheme

over 14 000 islanding/non-islanding cases are tested

considering different load active/reactive power values

load switching transients

capacitor bank switching

fault conditions in the main grid

different load quality factors

SNR levels

changes in network topology

and both types of conventional and inverter-based sources.