Proposal Title: Application of Innovation Statistics to Diagnose Biases in the HAFS system
Principal Investigator: Ryan Torn (U Albany, SUNY)
Collaborators:
NCEP/EMC
AOML/HRD
ABSTRACT:
Numerical weather prediction models are complex sets of equations designed to represent or parameterize numerous processes that dictate the evolution of the atmosphere. In the Hurricane Analysis and Forecasting System (HAFS), these equations are used to predict the evolution of tropical cyclone (TC) position, intensity, and eventually hazards, such as precipitation and wind. These models frequently contain regional or process-based biases, which in turn can yield biases in TC-related predictions. As a consequence, it is important to identify and alleviate the source of the biases, which can be difficult due to the complex interplay of different model components. This proposal, which primarily addresses Priority 4: “Develop improvements for tropical cyclone predictions, especially reducing hurricane intensity errors and improving the predictions for rapid changes in intensity”, aims to use statistics of the difference between short-term model forecasts and observations (i.e., innovation statistics) to identify and alleviate the source of model biases. In this work, we will develop a software framework that compares sets of short-term HAFS forecasts against a variety of observations that would allow the user to identify biases related to the large scale environment and TC vortex. Large-scale biases will be diagnosed by comparing HAFS forecasts against regular observations, such as rawinsondes or satellite winds, in an earth-relative framework, which can permit the identification of regional and/or phenomenological biases that can impact TC track or the environment. For TC intensity, we plan to compare HAFS forecasts against dropwindsondes, radar data, and other TC-specific observations. These comparisons will be done in a storm and shear-relative framework, which will facilitate computing statistics over many storms. Averaging the bias over many cases will identify where the model issues lie and suggest where improvements should be made and help make better use of observations within a data assimilation system. The results of these calculations will be communicated to the HAFS model developers, which will help them to identify priority areas for HAFS development. Finally, the techniques developed here can be extended to identifying biases in other UFS applications; therefore, the usefulness of this research can extend beyond HAFS. At the end of the project, we plan to make the software package available to the broader HAFS and UFS communities.
