Transient stability assessment datasets (TSA) are often imbalanced, a characteristic that negatively affects the performance of machine learning classifiers. In this work, a novel data-level method combining ADCHSMOTE-TL and lifting dimension linear regression is proposed to restore balance in imbalanced datasets of TSA. It consists of three major components: 1) an oversampling method based on convex hull theory; 2) a way to eliminate the generated samples of the non-target class using the Tomek links technique; and 3) a data-driven approach for efficient calculation of power flow equations. An essential advantage of the method proposed over the existing oversampling techniques is that it considers the nonlinear coupling between features in the TSA data. Case studies on the IEEE39 system have demonstrated that the proposed method can enhance diversity in sample generation, decrease the generation of non-target class samples, and improve the accuracy of the assessment model in detecting samples of transient instability.

class imbalance, machine learning, oversampling, power system, transient stability assessmen