Estimating the added value of GRACE total water storage and uncertainty quantification in seasonal streamflow forecasting

Abstract

Seasonal streamflow prediction is vital for water resources management and disaster mitigation. Multiple linear regression (MLR), as the simplest yet effective statistical model, is widely used in hydrological forecasting. To investigate the added value of incorporating GRACE total water storage anomalies (TWSA) in seasonal streamflow prediction and quantifying the uncertainty in model residuals, this study constructed four MLR models, namely a benchmark model, TWSA-incorporated model, uncertainty quantified model and synthetic model, to predict monthly streamflow in Yarlung Zangbo River, Upper Jinshajiang River, Xiangjiang River and Lanjiang River with lead times from 1 to 11 months. The results show that MLR models perform fairly well, with correlation coefficients greater than 0.8 and Nash-Sutcliffe efficiency coefficients up to 0.73. TWSA contributes moderately to forecast skill, whilst about 15% of the forecast skill is attributed to uncertainty quantification. The synergy of TWSA and uncertainty quantification brings basin-varying contributions, ranging from 5% to 26%.