Abstract. Seasonal predictions of river flow can be exploited among others
to optimise hydropower energy generation, navigability of rivers and
irrigation management to decrease crop yield losses. This paper is the first
of two papers dealing with a physical model-based system built to produce
probabilistic seasonal hydrological forecasts, applied here to Europe. This
paper presents the development of the system and the evaluation of its skill.
The variable infiltration capacity (VIC) hydrological model is forced with
bias-corrected output of ECMWF's seasonal forecast system 4. For the
assessment of skill, we analysed hindcasts (1981–2010) against a reference
run, in which VIC was forced by gridded meteorological observations. The
reference run was also used to generate initial hydrological conditions for
the hindcasts. The skill in run-off and discharge hindcasts is analysed with monthly
temporal resolution, up to 7 months of lead time, for the entire annual
cycle. Using the reference run output as pseudo-observations and taking the
correlation coefficient as metric, hot spots of significant theoretical
skill in discharge and run-off were identified in Fennoscandia (from January
to October), the southern part of the Mediterranean (from June to August),
Poland, northern Germany, Romania and Bulgaria (mainly from November to
January), western France (from December to May) and the eastern side of
Great Britain (January to April). Generally, the skill decreases with
increasing lead time, except in spring in regions with snow-rich winters. In
some areas some skill persists even at the longest lead times (7 months). Theoretical skill was compared to actual skill as determined with real
discharge observations from 747 stations. Actual skill is generally
substantially less than theoretical skill. This effect is stronger for small
basins than for large basins. Qualitatively, the use of different skill metrics
(correlation coefficient; relative operating characteristics, ROC, area; and ranked probability skill score, RPSS) leads
to broadly similar spatio-temporal patterns of skill, but the level of skill
decreases, and the area of skill shrinks, in the following order:
correlation coefficient; ROC area below-normal (BN) tercile; ROC area
above-normal (AN) tercile; ranked probability skill score; and, finally, ROC near-normal
(NN) tercile.