In order to describe the nonlinear relationship between the cavity backwater depth, cavity length and many influencing factors in aeration erosion reduction projects, and to achieve the accurate calculation of backwater and cavity length, a support vector machine regression model (GS-SVR) by means of machine learning was established based on the collection of 162 sets of model experimental data. Through the method of grid search and cross-validation, the relationship between the hyperparameters ( C and γ) in the support vector machine model and its influence mechanism on the accuracy of the model’s prediction were studied. On this basis, the model performance of six different input combinations (a total of twelve groups) was analyzed. The results show that the region with the best model performance can be found through grid search. In this region, C and γ show the opposite growth trend. Based on the above method, the optimal variable combination can be found ( i 1, i 2, Fr , Δ/h ) and the accurate prediction of the cavity length and the backwater depth is realized.