Scenario-driven probabilistic streamflow forecast based on the conditional vine copula and denoising diffusion probabilistic model

Study region This study focuses on two lakes in Northern China: Hongze (HZ) Lake and Luoma (LM) Lake. HZ Lake, the fourth largest freshwater lake in China, lies in the lower reaches of the Huai River and has a drainage area of 2596 km². LM Lake is situated in the Yi-Shu-Si River basin and covers a drainage area of 375 km². Study focus We proposed a novel scenario-driven probabilistic forecasting framework that integrates conditional vine copula (CVC) model with UNet-based Conditional Enhanced Denoising Diffusion Probabilistic Model (U-CEDDPM). The framework employed CVC to capture streamflow-teleconnection dependencies for next-month and water-year dry/wet classification forecasting while dynamically conditioning U-CEDDPM to generate high-fidelity probabilistic streamflow scenarios. New hydrological insights for the region Validation at HZ Lake and LM Lake in Northern China demonstrates that teleconnection-enhanced CVC improves water-year classification precision by 20.19 % at HZ Lake with significant gains in the detection of extreme dry, wet and normal events, while yielding marginal improvements at LM Lake. U-CEDDPM demonstrates superior deterministic performance, achieving 0.72 Nash Sutcliffe Efficiency (NSE) and 654.22 m³ /s root mean square error (RMSE) at HZ Lake, and 0.67 NSE with 157.49 m³ /s RMSE at LM Lake. Probabilistic evaluation also confirms U-CEDDPM's exceptional reliability with 66.28 % and 68.75 % prediction interval coverage probability at the 95 % confidence level for both lakes respectively, with competitive interval widths, significantly outperforming benchmark models (CVAE, CGAN, CDDPM).