Application of an Adaptive Post-Processor to Model Output Statistics with a View Towards Extension to the National Blend of Models

In this talk, we investigate the utility of extending the Roebber (2021) adaptive Artificial Neural Network (ANN) approach to “real world” NWP models by exploring potential applications to an updatable GFS MOS system for weather elements needed in support of the National Blend of Models (NBM) and LAMP. This talk details the proof-of-concept currently underway at NOAA’s Meteorological Development Laboratory, where the adaptive technique, originally applied to synthetic data from a complex version of a theoretical Lorenz circulation model, is modified for application to “simple” weather predictands (i.e. 3-hourly temperature and dewpoint, and day/night maximum-minimum temperature) in a manner analogous to MOS. The experiments are performed across time boundaries of previously known GFS version transitions since these changes can be “problematic” to the fidelity of the MOS guidance as currently structured. The adaptive technique is also developed and tested in a manner designed to seamlessly handle seasonal transitions, unlike the current seasonally stratified MOS forecast equations. It is expected that further refinement of the adaptive post-processing undertaken as part of this work will lead to approaches that are extensible to other aspects of NBM postprocessing.