Modelling uncertainties

Feigl et al. (2022) present a novel method of analyzing errors of process-based models attributing the model errors at each time step to specific input variables and model settings. This approach is helping to understand where model processes might need improvement, model input data might be of low quality or where model processes might be missing. The presented approach is a novel combination of (a) Machine Learning (using a data-driven model to learn predicting model errors), (b) Shapley Additive exPlanations and Principal Component Analysis (attributing errors to model inputs and variables), and (c) clustering (deriving groups of time steps that show similar error generation characteristics). The methodology is applied to the water temperature model HFLUX for a 3.45 km2 Canadian catchment. The results show that errors can be clustered in three groups related to specific processes indicating where model adjustments can lead to improved performance.
Moraga et al. (2022) present a new framework to quantify and partition the uncertainty in hydrological projections originating from climate models and natural climate variability. The approach is tested in the 478 km2 Kleine Emme and the 1730 km2Thur mountainous catchments in Switzerland. The study uses one emission scenario and nine climate models. The outputs of the climate models are stochastically downscaled using a two-dimensional weather generator producing a 90-member ensemble covering the period 2010-2089, and the hydrology is simulated using the spatially distributed TOPKAPI-ETH model. The results show that uncertainty of the annual streamflow projections is dominated by stochastic uncertainty due to large natural variability of precipitation. The same applies to extreme high flows. In contrary, snowmelt and liquid precipitation exhibit robust climate signals illustrating that streamflow uncertainty during warm seasons and at high altitudes are dominated by climate model uncertainty.