There are deep reasons why rare and severe events should be forecast in probabilistic terms [ Murphy, 1991]. They apply especially to precipitation, which is the major source of uncertainty in flood forecasts. And because precipitation amount is one of the most difficult meteorologic predictands, the uncertainty associated with any single estimate is high and varying from occasion to occasion. To allow the forecaster to convey the degree of uncertainty, to aid him in coupling principles of weather forecasting with principles of probabilistic reasoning which leads to the quantification of uncertainty, to formulate a scheme for judgmental processing of information from multiple sources, and to automate all algorithmic processing tasks are the objectives of a new methodology [ Krzysztofowicz et al., 1993].
The predictands are continuous variates: (1) the basin average precipitation amount accumulated during a period, and (2) a vector of fractions, which are the amounts accumulated during subperiods divided by the total amount for the period. Fractions, whose sample space is the simplex, are assumed to be stochastically independent of the total amount, the property verified statistically for river basins in Pennsylvania. A compromise between theoretical and operational considerations dictated the dichotomous forecast format: a probabilistic part comprises the exceedance function of the total basin average precipitation amount for the period; a deterministic part comprises the vector of expected fractions which specifies the expected temporal disaggregation of any total amount into subperiods.
The implementation of the methodology takes the form of a human-computer forecasting system. Its main components are judgmental tools, which guide the forecaster's reasoning, and estimation procedures, which elicit judgmental estimates from the forecaster and then transform this input into the final forecasts for all river basins within the service area. The input that the forecaster provides is in a graphical/alphanumerical format and consists of isopleths of the spatially averaged precipitation field with three exceedance fractiles assigned to each isopleth, and disaggregation zones, with the vector of expected fractions assigned to each zone.
The methodology has been tested operationally by the NWS
office in Pittsburgh, Pennsylvania, since August 1990.
Probabilistic QPFs are prepared twice daily, for the period of 24
hours disaggregated into four 6-hourly subperiods, and are
verified for two river basins of 5,786 km
and 3,389
km
in size. Through April 1994, over 3,500 forecasts have
been produced. Preliminary verifications of 1,005 forecasts
produced during the first year of testing indicate that
judgmental probabilistic QPFs for river basins have the potential
for being well calibrated, informative (relative to climatic
forecasts), and valuable for river forecasting [ Krzysztofowicz
and Drake, 1992].