In order to address the overfitting issue caused by the high level of sample redundancy in machine learning methods for transient evaluation of power systems

a sample selection method based on spatial distribution characteristics is proposed. Firstly

samples are categorized into boundary samples

edge samples

and internal samples according to their spatial distribution characteristics

and the impact of different types of samples on the model is analyzed. Secondly

the samples are pre-trained

the information content of the samples is sorted based on conditional information entropy

and the part with the largest information content is selected as the boundary sample. The Mahalanobis distance is then employed to eliminate sample variance and obtain geometrically positioned edge samples using a distribution elimination algorithm. By treating internal sample selection as a K-center problem

an optimal solution is obtained through utilization of the K-center greedy algorithm to acquire sparse representative internal samples. Finally

integration of all three types of extracted samples takes place. Simulation tests are conducted on both IEEE39-node system and East China Power grid system to validate the effectiveness of this proposed method.