Forced Oscillation Recognition with Interpretable Network Framework for Class Imbalance and Limited Real-World Data

Deep learning (DL) holds significant potential for distinguishing forced oscillations (FOs) from natural oscillations (NOs) and resonances in power systems. However, most existing studies treat DL models as black-box classifiers, and these models are often susceptible to challenges such as class imbalance among the three oscillation patterns (i.e., FOs, NOs, and resonances) and the limited size of real-world training datasets. These issues can lead to incorrect diagnoses of oscillation patterns in practical applications. To address these challenges, we propose a FO oscillation recognition framework that integrates balanced data processing, an interpretable recognition network, and a fine-tuning learning strategy. In this framework, DeepSMOTE1D is introduced to address the issue of class imbalance by oversampling features extracted from an encoder-decoder network. The interpretable DWT-CNN-LSTM recognition network incorporates multiple discrete wavelet transforms (DWT) into each convolutional and LSTM layer to enhance both interpretability and performance of the model. Finally, a fine-tuning transfer learning strategy is employed, where the recognition network is first pretrained on simulation data and subsequently retrained on limited real-world data. The effectiveness of DeepSMOTE1D and the fine-tuning strategy in the proposed framework is also evaluated. Experimental results show that the proposed recognition model not only learns effectively from simulated oscillation data but also outperforms existing related networks in real-world scenarios, achieving superior accuracy, precision, recall, and F1-score.