Unknown power quality disturbances classification based on transfer learning approach with imbalanced data considerations

• The proposed PQD model significantly reduces the computational burden and enhances robustness to new disturbances, ensuring high accuracy in real-time detection and classification. • The proposed model utilizes TL to enable efficient updates with a small dataset, improving adaptability in dynamic environments and making it suitable for online applications. • ADASYN is integrated with DL frameworks to address class imbalance by generating new samples for underrepresented classes, eliminating the need for Time-series simulations in real power systems. • WD is utilized to track distributional shifts in data, enabling timely model updates and ensuring adaptive performance in response to new disturbances. The integration of various new energy sources and power electronic devices into the power grid system has led to an increase in power quality disturbances (PQDs), necessitating accurate and efficient monitoring. However, maintaining model robustness against unknown disturbances and handling dataset imbalances remains a significant challenge in PQDs classification. In practical scenarios, disturbance occurrences are inherently unpredictable, and the operational conditions of the power system further contribute to their uneven distribution. This imbalance negatively impacts model efficiency. While a well-trained model can achieve high accuracy for a specific set of disturbances, its performance often deteriorates when exposed to unknown ones, necessitating continuous updates. This study addresses these two challenges with the help of transfer learning (TL) and adaptive synthetic sampling (ADASYN). To address dataset imbalance, ADASYN is employed to generate synthetic samples for minority classes, ensuring a more representative training distribution. Leveraging these balanced datasets, a convolutional neural network-long short-term memory (CNN-LSTM)-based model is trained to improve classification accuracy. Beyond data balancing, model uncertainty must also be quantified to determine when updates are necessary. For this purpose, Wasserstein distance (WD) is utilized as an uncertainty metric, where values exceeding 2.65 indicate high uncertainty, signaling the need for model updates. TL is employed to facilitate these updates with a small dataset of unknown samples, significantly reducing the computational burden for online classification. Evaluation results confirm the effectiveness of the proposed approach in PQDs classification.