Abstract: Calculation of natural or unimpaired flows is a key task in regional water resources assessment and hydroelectric project design. This paper presents an improved multiple fuzzy mean generating function model to forecast natural flows, based on several methods developed in previous studies. The model integrates this special function with the backstepping approach, extracting predominant periods and external factors, and optimal subset regression, and hence it is able to solve successfully the two frequently encountered problems in the conventional approaches: that they cannot effectively use the last data in a series, and that they produce results only statistically significant but inaccurate. We have verified this new model by testing its application to the natural runoff series at the Xianyang gauge station on the Wei River and examined the characteristics of probability distribution, statistical parameters, and precipitation-runoff relationship. Results show that it is an effective and rather accurate model.