As a typical air gap, the sphere gap is a kind of important electrode structure for studying air gap breakdown characteristics. Firstly, an electric field feature set is defined on the basis of electric field distribution on the shortest path. 271 power frequency breakdown voltages collected from IEC 60052: 2002 are considered as samples set. Then, four feature selection methods including the Pearson correlation coefficient, sensitivity coefficient, genetic algorithm and random forest are adopted to select features. In the end, the support vector regression (SVR) is trained by small sample data to predict power frequency breakdown voltages of other samples. The prediction results show that without the feature selection, the mean absolute percentage error is 4.21% for breakdown voltages of all samples with test results as the reference. Thus, the effectiveness of the feature set is verified. The mean absolute percentage errors of the results by adopting four feature selection methods are 1.92%, 2.00%, 3.86% and 2.04% respectively. Pearson correlation coefficient corresponds to the highest prediction accuracy. By comparing the results, it is concluded that the features of Emin and Lw90 play crucial roles in the model. The proposed method has a guiding significance to the study on discharge characteristics of complex gaps and engineering gaps in the future.