A vine copula-based polynomial chaos framework for improving multi-model hydroclimatic projections

Physically based hydrologic models have been extensively used for hydroclimatic projections, but key challenges remain owing to the heavy computational burden and structural variability of physically based models. In this study, we develop a vine copula-based polynomial chaos framework for improving multi-model projections of hydroclimatic regimes at a convection-permitting scale over the Dongjiang River Basin located in South China. Specifically, a deep neural network (DNN)-based polynomial chaos expansion (PCE) is developed to significantly improve the efficiency of probabilistic hydrologic predictions. A vine copula multi-model ensemble approach is also proposed to robustly combine hydrologic predictions generated from multiple DNN-based PCEs to improve reliability and accuracy. To assess regional hydrologic responses to changing climate, multi-decadal nested-grid climate projections over the Guangdong-Hong Kong-Macao Greater Bay Area (GBA) are developed using the convection-permitting Weather Research and Forecasting (WRF) model with 4-km horizontal grid spacing. Our findings reveal that the DNN-based PCEs achieve comparable performance to the physically based hydrologic predictions with an extremely low computational cost. The vine copula multi-model ensemble approach outperforms the Bayesian model averaging (BMA) by generating more accurate and reliable hydrologic predictions. The developed framework and physical models also lead to consistent projections of future changes in streamflow regimes. Our findings reveal that the projected increases in the frequency and intensity of extreme precipitation can lead to substantial increases in flood magnitudes, but the increases may not be obvious for river basins affected by multiple reservoirs.