Probabilistic streamflow prediction and uncertainty estimation in ephemeral catchments

Abstract

Probabilistic streamflow predictions at the daily scale are of major practical interest for environmental management and planning, including risk assessment as part of reservoir management operations. Ephemeral catchments, where streamflow is frequently zero or negligible, pose particularly stark challenges in this context, due to asymmetry of the error distribution and the discrete (rather than continuous) nature of zero flows. In this work, our focus is on two practical error modelling approaches where predictive uncertainty is approximated by a (transformed) Gaussian error model. The first approach, termed "pragmatic", does not distinguish between zero and positive flows during calibration, but sets negative flows to zero when making predictions. The second approach, termed "explicit", applies a "censored" Gaussian assumption in both calibration and prediction. We report a comparison of these two approaches over 74 Australian catchments with diverse hydroclimatology, using multiple performance metrics. The performance of the approaches depended on the catchment type as follows: (1) "mid-ephemeral" catchments, where 5-50% of days have zero flows, are best modelled using the "explicit" approach in combination with the Box-Cox streamflow transformation with a power parameter of 0.2; (2) "low-ephemeral" catchments, with fewer than 5% zero flow days, can be modelled using the pragmatic approach with (relatively) little loss of predictive performance; (3) "high-ephemeral" catchments, with more than 50% zero flow days, prove challenging to both approaches, and require more specialised techniques. The findings provide practical guidance towards improving probabilistic streamflow predictions in ephemeral catchments. Previous chapter Next chapter