A dual-stage self-attention hybrid convolutional neural network-gated recurrent unit architecture for real-time power quality disturbance diagnosis

The increased adoption of the smart grid (SG) infrastructure to overcome the limitations of traditional grid have faced growing challenges for real-time monitoring and classification of features due to the complexity and dynamic nature of power quality disturbances (PQDs). In particular, when PQDs overlap each other, traditional approaches relying on signal processing techniques tend to lose the dynamic, non-periodic characteristics which define certain PQDs. In this proposed work, a novel dual-stage hybrid architecture based on convolutional neural networks (CNNs) and gated recurrent units (GRUs) augmented by self-attention at both stages for real-time high performance PQD classification is presented. This novel hybrid architecture is allows for spatial and temporal feature extraction from electrical power signals, overcoming the weaknesses encountered in traditional approaches. In contrast to previous models that consider either local or sequential patterns independently, the proposed architecture combines both techniques enabling a more robust and computationally-efficient architecture demonstrates software-level real-time inference (21 ms) and provides a computationally efficient architecture suitable for future embedded deployment. The experimental outcomes on the benchmark databases compared to the state-of-the-art convolutional neural network-long short-term memory (CNN-LSTM) architecture offers improved execution time of 21 ms and a classification accuracy of 98.6%. It also outperforms both traditional CNN architecture with improved latency and high accuracy proposing it as a viable alternative for smart grid use cases like phasor measurement units (PMUs), microcontroller-based devices, etc.