From gridded runoff to streamflow: application of statistical post-processing to generate seasonal streamflow forecasts

Abstract. Hydrological models can be categorized as either fully distributed grid-based models or catchment-based models coincident with river gauging stations. Grid-based models provide national coverage at a relatively high spatial resolution. Catchment-based models target specific catchments upstream of a river-gauging station. While not providing seamless national coverage, catchment-specific model calibration allows catchment-based models to achieve improved performance relative to a distributed model using a single optimized parameter set. Catchment-based models often rely on the supply of real-time hydrological observations, the latency of which hinders the efficiency and timelessness of forecast generation and publication. This study evaluates whether a nation-wide grid-based hydrological model, coupled with statistical post-processing, generates comparable forecast skill to that of a catchment-based statistical hydrological model.

Two hydrological models are evaluated at 449 gauging stations across Australia. The Australian Water Resource and Assessment model (AWRA-L) is a 5km2 distributed landscape model that provides seasonal forecasts of key hydrological variables (soil-moisture and runoff) at the monthly temporal resolution. The necessary climate forcing is provided by the ACCESS-S seasonal climate forecasting model. Coupled with forecast post-processing, seasonal forecasts of river discharge (streamflow) can be generated from AWRA-L seasonal forecasts at gauged catchments across Australia. The second model evaluated in this study is a statistical hydrological model that is calibrated locally at the same individual catchments, where observed catchment conditions (antecedent observed streamflow) is the primary input variable with which to forecast streamflow.

Results from this study indicate that seasonal forecasts from a distributed hydrological model coupled with statistical post-processing achieve similar forecast skill to a locally calibrated statistical hydrological model where the observed catchment conditions represented by observed antecedent streamflow is the only input variable. The statistical post-processor is calibrated against historical streamflow observations but excludes antecedent observed streamflow as an input variable. For many catchments, AWRA-L can represent initial catchment conditions as well as the hydrological response arising from climate forcing from a seasonal climate model. The post-processing model provides a bias-correction that is a function of the magnitude of a hydrological model response only (e.g. runoff).

Forecast performance of the AWRA-L model is highest in the high-flow season across Australia but is limited in the low-flow season in certain regions (e.g. the tropical dry season in northern Australia). Inclusion of antecedent streamflow observations as an input variable results in excellent performance due to strong hydrological persistence in low flows. Depending on the location and time of year, either root-zone soil-moisture or runoff are found to maximise forecast skill. Root-zone soil-moisture maximises forecast skill for a higher proportion of sites than runoff overall, particularly in the low-flow season. As forecast skill increases, the proportion of sites for which runoff maximises forecast skill increases, particularly in the high-flow season. Results also indicate the degree to which climate forcing from a seasonal climate model contributes to forecast skill, as distinct to initial catchment conditions and hydrological persistence.