全球变化知识资源中心

Streamflow simulation is of great importance for water engineering design and water resource management. Most existing models simulate streamflow by establishing a quantitative relationship among climate, human activities, and streamflow and assuming the relationship is stationary in the long term. However, in a changing environment, this relationship may vary over time, resulting in the poor performance of many existing streamflow simulation models. In this study, inspired by data stream mining, adapting the gradient-boosted regression tree (XGBoost) to work in an online setting, a new statistically based model, called an online gradient-boosted regression tree (online XGBoost), is proposed to simulate streamflow dynamically in a changing environment. Here, the data of streamflow, climatic variables, and human activities are regarded as a data stream and the change in their relationships is treated as concept drift. The proposed model has two attractive properties. First, it makes it possible to capture the changed relationship between streamflow and its impact factors with the concept of a drift detection algorithm. Second, it can be used to simulate streamflow dynamically by updating models based on the concept drift detection results. Taking the Qingliu River catchment as a case study, the results show that the proposed method achieved good performance on monthly streamflow simulations during 1989 and 2010 with a Nash-Sutcliffe model efficiency coefficient (NSE) of 0.73. Furthermore, it outperformed comparable methods, including four statistically based methods (online support vector regression, online regression tree, online random forest regression, and online boosting tree regression) and four lumped parameter hydrological models (SimHyd, Sacramento, soil moisture accounting and routing, and Tank). The proposed model provides a useful tool for streamflow simulation in a changing environment. Findings will help water resource managers adapt to climate change.