An oil production prediction method based on multi-model combination under Stacking ensemble learning was proposed associated with the frontier theory of machine learning. The model was used to predict the dynamic oil production from the production data of a domestic ultra-high permeability oilfield in China developed by water flooding. Considering the differences in training principles of different algorithms, the XGBoost algorithm, long and short memory network (LSTM), temporal convolutional network (TCN) and other models are selected as base learners, the MLR algorithm is chosen as meta learner. The results show that the Stacking ensemble model has smaller average error and better prediction robustness compared with the traditional single model, since the ensemble model fully combined the advantages of each base learner. The proposed model is of great significance to the application in ultra-high water cut reservoirs.