Two-stage weakly supervised learning to mitigate label noise for intelligent identification of power system dominant instability mode

Deep learning (DL) is effective in identifying the dominant instability mode (DIM) of power systems. However, regular supervised learning for training DL models requires a large number of training samples with accurate labels provided by power system experts, which is prohibitively costly in practice. To address this issue, this paper proposes a weakly supervised learning framework to train DL models for DIM identification based on cheap but potentially inaccurate (noisy) labels from non-expert engineers. The framework comprises two stages to mitigate the detrimental effects of label noise. At Stage I, an auxiliary model is proposed to intelligently detect detrimental noisy samples while preserving truly-labeled informative hard samples based on the entire training loss dynamics of base DL models. Then Stage II incorporates virtual adversarial training to utilize all samples, including noisy ones with labels removed, to train a smooth DIM identification model in a semi-supervised learning way. It can help further mitigate the effects of undetected label noise. Case studies are conducted on CEPRI 36-bus system and Northeast China Power System (2131 buses). The results verify that the proposed framework can tolerate high-intensity feature-(in)dependent label noise and build reliable DL models for DIM identification with significantly less reliance on experts.