Using Evolutionary Programming to Add Deterministic and Probabilistic Skill to Spatial Model Forecasts

Evolutionary programming is applied to the postpocessing of ensemble forecasts of temperature on a spatial domain. These forecasts are obtained from the 11-member Reforecast V2 ensemble over the region from 24°–53°N to 125°–66°W for the period 1 January 1985–14 May 2011. The evolution is based upon a static ecosystem model that holds constant the number of individuals (algorithms), using a fixed rate of introduction of new algorithms and removal of existing algorithms. Each algorithm adheres to a specific underlying genetic architecture, and the selection pressure on the “species” is according to deterministic performance (root-mean-square error) on a training dataset. On a 2325-case, independent test dataset, the method improved root-mean-square error and ranked probability score relative to the Reforecast ensemble by 0.31°F (8.7%) and 3.3%, respectively, across the domain, with 96% of the grid points showing simultaneous improvements in both measures. The use of input information by the evolutionary programming algorithms varied by region; while the algorithm forecasts at all locations are fundamentally tied to the Reforecast ensemble forecast, northeastern locations found snow cover to be the next most useful input, whereas southwestern locations preferentially employed precipitable water. An adaptive form of the approach, developed to be readily implemented into operations, is tested in the absence of improving inputs but is found to only slightly degrade performance (1.2% in root-mean-square error and 0.6% in ranked probability skill score). A number of future extensions are discussed.