Dynamic Imbalance Correction for Transient Stability Assessment Based on Interval Sample Impact Norm

The robustness of real power systems results in extremely rare transient instability events, which inevitably causes severe sample imbalance issues in data-driven transient stability assessment (TSA). Training TSA models on such imbalanced datasets typically introduces significant assessment bias, which severely degrades model reliability. In existing solutions, samples quantity imbalance is considered, but quality imbalance is ignored. In response to this limitation, a dynamic imbalance correction method considering both quantity and quality imbalance is proposed. Firstly, samples exhibiting similar impacts are systematically categorized into different intervals based on the probability distribution of fault severity. Subsequently, the imbalance degree in each interval is accurately quantified by sample impact norm mean ratios. A cost-sensitive matrix is constructed based on the derived imbalance degree and dynamically adjusted during training process. Finally, the loss function is modified by using the cost-sensitive matrix to alleviate the assessment bias. The simulation results in the IEEE-39 bus system and the East China power grid system show that the corrected model achieves accuracies of 98.41% and 97.80% in the two power systems, respectively. Meanwhile, the proposed method is also applicable to other deep learning models with different structures. Compared with the traditional cost-sensitive method, it has better correction performance, and the training time is shortened by 38.81% and 46.79% in two system, while achieving better convergence effect and continuous stability update.