训练集再利用：一种基于相似样本的电力系统主导失稳模式识别物理可信框架

Accurate and physically reliable identification of the dominant instability mode (DIM) is crucial for ensuring secure and stable power system operation. Data-driven models, particularly deep learning (DL), have achieved significant progress in addressing this challenge. However, the “black-box” nature of DL limits interpretability, leading to unreliable outcomes that conflict with the stringent reliability requirements of power systems. To solve this, a novel DIM identification framework is proposed to improve accuracy and reliability by re-applying training set samples. First, a training method is proposed to enhance the noise robustness and ability to cluster similar samples of the DIM model, achieving high-accuracy DIM identification. Furthermore, a two-stage interpretability algorithm is developed. In Stage 1, a fixed-radius k-nearest neighbor (KNN) algorithm matches the low-dimensional features of the DIM model for the target sample to find similar samples. In Stage 2, the knowledge-embedded Fréchet distance analyzes the differences between the target sample and similar training samples, using the physical discriminative logic of the most similar sample to guide predictions and assess reliability. The effectiveness of the proposed framework is validated on the CEPRI 36-bus power system and the Northeast China Power System (2131 buses), demonstrating improvements in both DIM prediction accuracy and physical reliability.