Air gap is the most commonly used insulation medium in ultra-high-voltage transmission and transformation projects. The reasonable configuration of air gap is very important for the safe and stable operation of power system. In this paper, based on 360 sets of rod-plane gap switching impulse discharge voltage data from 9 test sites, the gap distance, air pressure, dry temperature, absolute humidity and test site are selected as input features, and the TabNet discharge voltage prediction model is established. The results of the test set show that the root mean square error of TabNet is 43.66 kV, and the mean absolute percentage error is 3.48%. In comparison with AdaBoost-SVR and LightGBM algorithms, the optimal prediction results are obtained by TabNet, which provides a more reliable intelligent algorithm for air gap discharge prediction.