Abstract: Based on the nature that river flow series is a kind of seasonal non-stationary time series, a proper combinatory model describing this nature is developed in this paper. The advantage of the combinatory model is that information in historical data can be fully utilized, i.e., the correlated information of time series sampled both yearly or monthly can all be used. FORTRAN programs have been written to forecast medium-long terms flow of river by means of this combinatory model. The monthly mean flow forecasted one to twelve months ahead of time at Hankou Station on the Long River and Longtan Station on the Hongshuhe River show that the forecasting effect is satisfactory. The work is part of a water resource research work mandated by the Ministry of Hydro- power.