Bias learning improves data driven models for streamflow prediction

Abstract

Study region: Andun river basin of Southern China and 273 watersheds across the continental United States. Study focus: Owing to the data incompleteness and the changing environment, there is bias existing in the data driven models for streamflow prediction, greatly limiting their application in actual practice. In this paper, we incorporate the bias learning components into the data driven models to establish the mapping-bias-learning models, and design two groups of experiments, one in the Andun River basin in China and the other in 273 watersheds in the continental United States (CONUS). In the first group of experiments, we respectively apply three machine learning algorithms and one traditional statistical method to generate sixteen mapping-bias-learning models ae well as four mapping-learning-alone models. We also explore the effectiveness of different bias learning strategies, including multiple bias learning, stacking bias learning and incremental updating bias learning. In the second group of experiments, we apply the mapping-bias-learning models to the streamflow prediction in 273 watersheds of CONUS to verify the universality and robustness of the bias learning method. New hydrological insights for the region: The mapping-bias-learning models significantly outperform the mapping-learning-alone models, and the machine learning methods are superior to the traditional statistical method in term of the ability of bias learning. In addition, adopting the appropriate bias learning strategies can further improve the streamflow forecast performance of data-driven models. The bias learning method is applicable to various watersheds with different hydrological conditions and play efficient roles at different streamflow levels. The median of MRE drops from 0.156 for mapping-learning-alone models to 0.131 for mapping-bias-learning models, while median of NSE increases from 0.918 for mapping-learning-alone models to 0.932 for mapping-bias-learning models. More inspiringly, the bias learning method can still play roles in data driven models under the environmental conditions that never occurred before, suggesting its remarkable superiority. This study highlights the power of the bias learning method for data driven models in streamflow prediction, and the promising prospect of the combination between bias learning and machine learning methods in hydrological modelling.

Keywords

• Streamflow prediction
• Data driven model
• Bias learning
• Machine learning