The invention discloses a precipitation prediction method based on a deep network. The precipitation prediction method comprises specific steps as follows: collecting hydrological environment data, preprocessing the data and normalizing the data in a 01 interval; detecting the sufficiency of environmental factors by using KMO test and Bartlett sphericity test, and screening the environmental factors with a factor analysis method after test; using a divided training sample as an input value of the deep network, subjecting the network to hierarchical unsupervised learning and then performing reverse fine adjustment on the whole network by use of a top BP network to guarantee the precision; using a test sample as the input of a trained model and inversely-normalizing the output obtained from the model to obtain the predicted result of precipitation. The precipitation of an area in the future can be predicted more accurately, the factors with lower correlation are screened out with the factor analysis method, the calculation efficiency is improved, and the scientificity and reasonability of the precipitation prediction process are greatly enhanced.