Abstract: The phenomenon that hydrological series often present nonstationary under changing environments brings huge challenges for the traditional frequency analysis methods based on the independent and identical distribution hypotheses. Thus, it is urgent to investigate new frequency analysis methods for the nonstationary hydrological series. In this paper, a nonstationary hydrological frequency analysis method considering local trend was proposed. The method consists of three main procedures. Firstly, a moving window trend analysis method with identifying capability both for local and global trend was employed. Secondly, time-varying distribution models were established and their parameters were estimated by generalized maximum likelihood method coupled with Bayesian theory. Finally, selecting the optimal model and then design values and confidence intervals were calculated according to the equal reliability principle. The proposed method was applied to the frequency analysis of the annual runoff series gauged at the Hongyashan Reservoir in the Shiyang River basin, Gansu province. Results show that from the perspective of design values and 95% confidence intervals, there are certain difference between the new method and the conventional stationary frequency analysis method, as well as the time-varying distribution method considering the overall trend. This paper reveals that the nonstationary hydrological frequency analysis on the basis of accurately identifying local or overall trend component of the sequence is conducive to estimating design values reasonably.