Abstract: Reliable reservoir inflow forecast is the basis of long-term reservoir scheduling decision. This paper describes a new method of rolling flow forecast that gradually increases forecast horizon using the artificial neural network to overcome the limitation of traditional medium- and long-term flow forecast models on forecast horizon and their uncertainty in flow forecasting. A Copula function is adopted to construct joint distribution functions for the sequences of prediction errors and realize random simulation of hydrological prediction errors, and thus this method can describe flow forecast uncertainties quantitatively. Calculations of the monthly average flow and the uncertainties for the Three Gorges reservoir in late non-flood season show that our new method is satisfactory and applicable to practical operation of rolling forecast. Copula functions well describe the correlation between prediction relative error sequences, and often produce small deviation of the correlation coefficients, statistics, and empirical distributions of the simulated sequences from those of the observed sequences.