Sie befinden Sich nicht im Netzwerk der Universität Paderborn. Der Zugriff auf elektronische Ressourcen ist gegebenenfalls nur via VPN oder Shibboleth (DFN-AAI) möglich. mehr Informationen...
Single process-based rainfall–runoff model can hardly capture all the runoff characteristics, especially for flood periods and dry periods. In order to address the issue, we have employed a new hydrologic model, XXT model, to be the base learner of AdaBoost.RT algorithm to make a robust aggregation, namely AdaBoost-XXT.
•AdaBoost.RT algorithm is firstly applied in a process-based rainfall–runoff model.•Four modifications are made based on the original version of AdaBoost.RT algorithm.•PSO algorithm is employed for the determination of model parameters.•A new ensemble model, namely AdaBoost-XXT, is initially presented in this paper.
High-efficiency rainfall–runoff forecast is extremely important for flood disaster warning. Single process-based rainfall–runoff model can hardly capture all the runoff characteristics, especially for flood periods and dry periods. In order to address the issue, an effective multi-model ensemble approach is urgently required. The Adaptive Boosting (AdaBoost) algorithm is one of the most robust ensemble learning methods. However, it has never been utilized for the efficiency improvement of process-based rainfall–runoff models.
Therefore AdaBoost.RT (Adaptive Boosting for Regression problems and “T” for a threshold demarcating the correct from the incorrect) algorithm, is innovatively proposed to make an aggregation (AdaBoost-XXT) of a process-based rainfall–runoff model called XXT (a hybrid of TOPMODEL and Xinanjing model). To adapt to hydrologic situation, some modifications were made in AdaBoost.RT. Firstly, weights of wrong predicted examples were made increased rather than unchangeable so that those “hard” samples could be highlighted. Then the stationary threshold to demarcate the correct from the incorrect was replaced with dynamic mean value of absolute errors. In addition, other two minor modifications were also made. Then particle swarm optimization (PSO) was employed to determine the model parameters. Finally, the applicability of AdaBoost-XXT was tested in Linyi watershed with large-scale and semi-arid conditions and in Youshuijie catchment with small-scale area and humid climate. The results show that modified AdaBoost.RT algorithm significantly improves the performance of XXT in daily runoff prediction, especially for the large-scale watershed or low runoff periods, in terms of Nash–Sutcliffe efficiency coefficients and coefficients of determination. Furthermore, the AdaBoost-XXT has the more satisfactory generalization ability in processing input data, especially in Linyi watershed. Thus the method of using this modified AdaBoost.RT to enhance model performance is promising and easily extended to other process-based rainfall–runoff models.