Abstract: Water yield is an important index to evaluate engineering quality for underground water-sealed rock cavern.The commonly used analysis prediction at precent is mainly based on equivalent continue media model, so its prediction value has a certain deviation against surveillance value, and not matching the engineering requirements. For the purpose of improving accuracy of prediction, based on analysis various factors impacting water yield, the nonlinear artificial neural network（ANN） model which was adopted to predict water yield of underground water-sealed rock cavern has been established, and real data from completed engineering in domestic was used as training dataset. The analysis prediction in this model shows that the ANN is a kind of simple, quick and convenient method with better accuracy and efficiency, which has good generalization performance, especially when dealing with some undetermined factors such as complicated theoretical model and geological parameters. It’s an efficient and practical method to predict water yield of underground water-sealed rock cavern.