Using a Self-Learning Algorithm for Single-Station Quantitative Precipitation Forecasting in Germany

Abstract

A self-teaming algorithm called goal-orientedpattern detection was used to develop a set of 12 models designed to forecast 24-h precipitation amounts for eight sites in southern Germany. The forecasts of expected precipitation amount valid for the following 24-h period are issued shortly after 0000 UTC each day and are based on the available rawinsonde data from the current 0000 UTC and previous 1200 UTC observations. The period 1973?1982 was used for developing the forecast models, and the year 1983 was used for verification purposes. The forecast models provide the probability of precipitation greater than any specified amount at each of the eight stations. The overall skill score (percentage reduction in the squared forecast error compared to climatology) for 1983 over five forecast amounts was 31%. The forecast skill for measurable precipitation was 37% and decreased with increasing precipitation amounts to 19% for amounts greater than or equal to 0.20 in. The forecast model executes on an MS-DOS-based personal computer and provides the probability of precipitation greater than any specified amount or specifies the amount of precipitation associated with any given risk level. These values can be shown in tabular form for each station or displayed as a contour map over the region of interest.