Interpretable probabilistic modeling method for runoff prediction: A case study in Yangtze River basin, China

Abstract

Study region: In the Yangtze River basin of China. Study focus: Accurate and reliable streamflow forecasting methods are of utmost importance for the sustainable utilization and management of water resources. Recently, machine learning models have been widely used in streamflow forecasting, but greatly focus on the improvement of prediction accuracy and ignore the description of prediction uncertainty. Therefore, we applied bidirectional gated recurrent unit (BIGRU), attention mechanism (AM), and natural gradient boosting (NGB) methods to predict the uncertainty of streamflow in the Yangtze River basin. We propose a novel hybrid streamflow probabilistic prediction model BIGRU-AM-NGB. In addition, the underlying contribution of input features to output was thoroughly analyzed. New hydrological insights for the region: Six state-of-the-art models were compared with the proposed model from three aspects: deterministic prediction, interval prediction, and probability prediction. The results revealed that BIGRU-AM-NGB model can obtain high accuracy point prediction, reliable probability prediction, and appropriate interval prediction. Different input features result in different contribution to output. The runoff of Yichang hydrological station one-day ahead dominated the output, followed by the runoff of Zhutuo hydrological station two-day ahead. This can be attributed to the spatial and temporal distribution of hydrographic stations. Therefore, BIGRU-AM-NGB can achieve high-precision streamflow predictions while quantifying uncertainty, and effectively develop implicit relationships between features and output that obey physical properties.

Keywords

• Probabilistic forecasting
• Interpretable machine learning
• Natural gradient boosting
• Bidirectional gated recurrent unit
• Forecast uncertainty
• Streamflow