Through analyzing problem of wavelet analysis, the annual precipitation series from 1956 to 2000 in the sub-water resources region of upper Lanzhou is decomposed into muti-time scale series with EMD method. The results show that the precipitation series has periods that about 3, 4--8, 11 years, and the physical backgrounds and trends of the IMF sub-series are discussed. An annual precipitation-runoff forecasting ANN model based on EMD is established, with the EMD decomposition series as input and the corresponding annual runoff as output. The study shows that, as a new and original signal decomposition method, Empirical Mode Decomposition namely EMD can be used as a tool to decompose hydrological time series into exact muti-time scale sub-series for finding their local change rule, and then to supply input variables with high quality and muti-level to enhance model quality.