Tuning AutoNowcaster automatically

Abstract

AutoNowcaster (ANC) is an automated system that nowcasts thunderstorms, including thunderstorm initiation. However, its parameters have to be tuned to regional environments, a process that is time consuming, labor intensive, and quite subjective. When the NationalWeather Service decided to explore using ANC in forecast operations, a faster, less labor-intensive, and objective mechanism to tune the parameters for all the forecast offices was sought. In this paper, a genetic algorithm approach to tuning ANC is described. The process consisted of choosing datasets, employing an objective forecast verification technique, and devising a fitness function. ANC was modified to create nowcasts offline using weights iteratively generated by the genetic algorithm. The weights were generated by probabilistically combining weights with good fitness, leading to better and better weights as the tuning process proceeded. The nowcasts created by ANC using the automatically determined weights are compared with the nowcasts created by ANC using weights that were the result of manual tuning. It is shown that nowcasts created using the automatically tuned weights are as skilled as the ones created through manual tuning. In addition, automated tuning can be done in a fraction of the time that it takes experts to analyze the data and tune the weights. © 2012 American Meteorological Society.