Continuous rainfall-runoff model comparison and short-term daily streamflow forecast skill evaluation

Select | Print

Pagano, Tom; Hapuarachchi, Prasantha; Wang, QJ

Pagano, Tom; Hapuarachchi, Prasantha; Wang, QJ




This research was done to support the development and improvement of an operational short-term (i.e. <14 days into the future) streamflow forecasting service at the Australian Bureau of Meteorology. Part of this research is to compare various available rainfall-runoff models and determine which model would be suitable for the task. Four rainfall runoff models (AWBM, GR4J, PDM, and SimHyd) were run on 240 Southeast Australian catchments from 1974-2006 and the results compared to two naïve baseline models (rainfall multiplied by a ratio and a simple loss model that has been used experimentally in the Bureau operational environment). For practically every measure, GR4J outperformed every other model considered. GR4J rose to the task of performing well in the measures it was asked to perform well in during model setup. Part of the data was withheld during model setup and GR4J had the least decline in performance of any of the models when asked to simulate streamflow from this independent period. Most remarkably, it also performs among the best of all the models in measures that were not part of the model setup (related to the shape and spikiness of the hydrograph). This suggests that GR4J is capturing realistic hydrologic behaviour and is not simply fitting the data. This is all the more remarkable because GR4J has 4 tunable parameters, fewer than most rainfall-runoff models. The models were tried in a variety of configurations and tested on the 240 catchments to determine the optimal setup. A tunable parameter to rescale rainfall inputs did not reliably improve the four rainfall runoff models. Additional catchment routing routine marginally benefitted SIMHYD (which already had ways of delaying and attenuating surface runoff), but it did not improve GR4J (which uses a unit hydrograph). A skill-weighted average of the hydrographs from the four models was not significantly better than GR4J on its own. Error correction uses the difference between recent observed and simulated flow to improve the simulations. An error correction routine that corrects for quickly (1-day) and slowly (1-year) varying biases was proposed and applied. This is a novel approach and showed the most improvement during low flow conditions. The naïve models benefitted the most, although the final performance of uncorrected rainfall-runoff models was still better than the corrected naïve model. The Ensemble Streamflow Prediction method was used to generate retrospective forecasts (as opposed to simulations where future rainfall is known with certainty). Streamflow forecast skill diminished with lead-time, sometimes rapidly (i.e. +1 day ahead). A fair amount of the apparent forecasting skill was due to the model’s ability to reproduce the seasonal cycle of streamflow. When compared to a more challenging baseline that considered the seasonal cycle, many catchments did not have forecast skill beyond +3 days ahead. To the authors’ knowledge, this is the first Australian study to measure short-term streamflow forecasting skill at more than 125 catchments.


WIRADA; 4.1 Water forecasting and prediction - short term

Surfacewater Hydrology

Manuscript (pdf) (6.10MB)

This report has been placed on the CSIRO repository and may be made available to persons outside of CSIRO for non commercial purposes, in its entirety and without deletion of disclaimers and copyright information.

Water for a Healthy Country National Research Flagship


Technical Report (Author)


Pagano, Tom; Hapuarachchi, Prasantha; Wang, QJ. Continuous rainfall-runoff model comparison and short-term daily streamflow forecast skill evaluation. CSIRO; 2010. csiro:EP103545.

Loading citation data...

Citation counts
(Requires subscription to view)