LAUSR

Abstract Due to the effects of frequent anthropogenic activities and climate change, the natural annual runoff series presents typical nonlinear, non-stationary and multiple-scale characteristics, which triggers the common problem of low accuracy for long-term runoff predictions. Therefore, the main goal of this study was to improve the long-term prediction accuracy for the nonlinear and non-stationary runoff series by introducing a new hybrid approach based on improved ensemble intrinsic time-scale decomposition (IEITD) and improved nearest neighbor bootstrapping regressive (INNBR) methods. First, the Brock-Dechert-Scheinkman (BDS) test and Augmented Dickey-Fuller (ADF) test were used to identify the nonlinear and non-stationary characteristics of annual runoff series, respectively, and the Modified Mann–Kendall (MM-K) test and Mann–Kendall (M-K) abrupt change test methods were applied to explore the drivers of non-stationarity. On this basis, the IEITD was used to decompose the original annual runoff series into several proper rotation components (PRCs) and a monotonic residual trend term to make the non-stationary runoff time series stationary. Then, the INNBR model was applied to predict the respective PRCs. The residual series was predicted by the polynomial fitting method. Finally, the predictive results of each PRC and residual series were summed to obtain an ensemble forecast for the runoff series. The performance of the new hybrid approach was tested by the annual (nine hydrological stations) and monthly (three hydrological stations) runoff data covering 1956–2011 in the Luanhe River basin in China. Results suggested: (1) the annual runoff time series of nine hydrological stations presented obviously dependent nonlinear structure; (2) the annual runoff series of nine hydrological stations were all non-stationary; (3) human activity, rather than change in precipitation, was the major driving factor of runoff decline in the Luanhe River basin; (4) For the performance evaluation criteria of Nash-Sutcliffe efficiency coefficient (NSEC), compared with Nearest Neighbor Bootstrapping Regressive (NNBR) and INNBR, the precision of IEITD-INNBR model almost increase by 227% and 37% on the average of nine hydrological stations; (5) the new hybrid approach of combining the IEITD and INNBR models outperformed the other two models (NNBR and INNBR) tested, and it is capable of capturing the nonlinear, non-stationary and multiple-scale characteristics of complex runoff time series and obtaining higher predictive precision.