A hybrid process-based machine-learning model for assessing future change in streamflow predictability in relation to snowpack change
Abstract
Climate induced snowpack loss is affecting the magnitude and timing of snowmelt driven spring runoff, with implications on the streamflow predictability, as it becomes more dependent on the unknown meteorological forecasts and less dependent on the known snowpack initial conditions. In this study, we use a hybrid-modelling framework consisting of projections from the Variable infiltration capacity (VIC) hydrologic model as pseudo-observations, and Bayesian regularized neural networks (BRNN) as an emulation model for evaluating future changes in streamflow predictability. We use the BRNN in an ensemble streamflow prediction approach, with inputs derived from the climatological traces of precipitation and temperature from a suite of GCMs and VIC simulated snow water equivalent (SWE). We apply the framework over two western Canadian River basins and assess future changes in warm season mean and maximum flow predictability in terms of a number of deterministic and categorical skill metrics. The results indicate contrasting patterns of change in the predictability skills, both across the two basins and streamflow variables. Specifically, under future climate scenarios, while the basin-scale mean flow predictability skills decline for two basins, the headwater subbasin-scale skills increase or decrease marginally. In the case of maximum flow, while the future predictability skills decline for the warmer basin, the skills improve across the colder basin and headwater subbasins. An explanation for the improved maximum flow predictability is provided by the reduced future forecast horizon due to its earlier timing, offsetting the effect of SWE loss. Further analyses revealed that the change in variable importance (VI) as the main determinant of the change in predictability, with reduction in VI of SWE resulting in lower predictability and visa versa. Overall, the results reinforce the flexibility and robustness of the hybrid modelling approach, with good emulation model fits and physically plausible change in streamflow prediction skills.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H35D1071S