Model averaging methods to merge operational statistical and dynamic seasonal streamflow forecasts in Australia. Issue 3 (31st March 2015)

• Record Type: Journal Article
• Authors: Schepen, Andrew; Wang, Q. J.
• Is Part Of: Water resources research. Volume 51:Issue 3(2015:Mar.)
• Page Start: 1797
• Page End: 1812
• Publication Date: 2015-03-31
• Subjects: Hydrology -- Periodicals; 333.91
• DOI: 10.1002/2014WR016163
• Languages: English
• ISSNs: 0043-1397

Abstract

<abstract abstract-type="main"> <title>Abstract</title> <p>The Australian Bureau of Meteorology produces statistical and dynamic seasonal streamflow forecasts. The statistical and dynamic forecasts are similarly reliable in ensemble spread; however, skill varies by catchment and season. Therefore, it may be possible to optimize forecasting skill by weighting and merging statistical and dynamic forecasts. Two model averaging methods are evaluated for merging forecasts for 12 locations. The first method, Bayesian model averaging (BMA), applies averaging to forecast probability densities (and thus cumulative probabilities) for a given forecast variable value. The second method, quantile model averaging (QMA), applies averaging to forecast variable values (quantiles) for a given cumulative probability (quantile fraction). BMA and QMA are found to perform similarly in terms of overall skill scores and reliability in ensemble spread. Both methods improve forecast skill across catchments and seasons. However, when both the statistical and dynamical forecasting approaches are skillful but produce, on special occasions, very different event forecasts, the BMA merged forecasts for these events can have unusually wide and bimodal distributions. In contrast, the distributions of the QMA merged forecasts for these events are narrower, unimodal and generally more smoothly shaped, and are potentially more easily communicated to and interpreted by the forecast users. Such special occasions are found to be rare. However, every forecast counts in an operational service, and therefore the occasional contrast in merged forecasts between the two methods may be more significant than the indifference shown by the overall skill and reliability performance.</p> </abstract>