|
A winning proposal for the Innovative Research Program, 2007:
Improving the Initialization of Hurricane Forecast Using an Ensemble-based Data Assimilation Technique
Investigators: Xuguang Wang*, Tom Hamill^, and Jeff Whitaker^
*CIRES Climate Diagnostics Center,
^NOAA Earth System Research Laboratory
Email: xuguang.wang@noaa.gov, Tel: 303-4974434
Objectives: For a case of a poorly forecast hurricane (Rita, in 2005), we propose a pilot study to (1) determine whether ensemble-based data assimilation (Ens-DA) techniques improve the quality of
initial conditions (analyses) and the subsequent intensity and track forecasts relative to those
initialized from the current standard, 3-dimensional variational (3D-Var) techniques, and (2)
determine whether utilizing an ensemble of multiple physical parameterization schemes in Ens-DA
to represent the uncertainty of the model will further improve the analyses and forecasts.
Background: Numerical predictions of hurricanes have improved greatly in the past few years
through the improvement of the numerical weather prediction models. Unfortunately, there has been
less progress in improving the initialization for these models, and thus the accuracy of the hurricane
predictions is still limited.
Data assimilation is a process of blending together prior short-term forecast(s) with new
observations. Error statistics are required for both the forecast and the observations in order to know
how to combine the two. The widely used 3D-Var data assimilation (DA) technique typically
utilizes a simple, unchanging model for the forecast error statistics, one that assumes, for example,
the forecast errors are the same in the eye wall of a hurricane as they would be 500 km away from the
eye. Such a forecast- error model provides a poor approximation to the complex, flow-dependent
error structure around hurricanes. In Ens-DA, parallel DA and forecast cycles are conducted, and
flow-dependent error statistics are estimated using ensemble forecasts. The improved covariance
model offered by the ensemble may result in a more appropriate adjustment of the forecast to the
observations, deeper and more realistic initial vortices, and improved hurricane forecasts. Figure 1
illustrates this with a synthetic hurricane ensemble. The synthetic ensemble was generated from a
time series of 50 hourly hurricane Katrina forecasts from the Weather Research and Forecasting
Model (WRF). These 50 forecasts were repositioned so that the eyes were co-located, though the
50 forecasts provided differing hurricane structures around that eye. The ensemble mean
(background forecast) of the 850 hPa hurricane wind is shown in panel (a). Assume we assimilate a
single southerly wind observation located at the black dot, and the wind here is stronger than the
background forecast. Panel (b) shows the adjustment of the background forecast to this observation
by the Ens-DA. The error model generated from the ensemble of hurricane simulations diagnoses a
correlation of wind strength around the eye wall. Consequently, the observation is able to coherently
strengthen the entire eyewall vortex. In contrast, the 3D-Var adjustment (panel c) is localized near
the observation site, thereby adding an unrealistic, asymmetric wind field to the forecast.
 Fig. 1. Illustration of the usefulness of the Ens-DA for tropical cyclones from a synthetic ensemble. (a) is the ensemble-mean 850 hPa wind. The arrows denote the wind vectors and the color shades denote the vector magnitudes (b) The Ens-DA analysis increment of 850 hPa wind from the assimilation of one 850 hPa V wind observation with 5ms-1 observation increment at the black dot. (c) As in (b), but for the 3D-Var.
In Ens-DA, the ensemble of short-term forecasts should produce a realistic diversity of
forecasts, capturing the error growth due both to initial-condition errors (chaos) and numerical model
imperfections. One source of model error is the representation of subgrid-scale physical processes
(physical parameterization). One way to represent such model error is to use different physical
parameterization schemes, which are conveniently available from the WRF model that we will use in
this work. We hypothesize that the utilization of multiple physical parameterization schemes (multiphysics)
in the ensemble will produce a more realistic representation of the distribution of possible
hurricane forecast states, which consequently will improve the adjustment to observations in the Ens-
DA.
Importance: The proposed pilot work may demonstrate a dramatically better hurricane initialization
and forecast method. Also, if hurricanes can be more appropriately modeled with Ens-DA
techniques, this may open a new frontier of research into understanding the causes of rapid hurricane
intensification; the diversity of hurricane simulations, with some strengthening and others
weakening, can be used to understand what are the dynamical characteristics required for intensity
changes.
What makes this innovative? Application of Ens-DA techniques to hurricanes is just beginning.
To our knowledge, nobody has yet tested Ens-DA in conjunction with a multi-physics ensemble.
Why is this interdisciplinary and extending the frontiers of science? The conduct of the project
requires expertise in both meteorology and statistical estimation theory. The methodology can be
extended to analyze tropical cyclogenesis, which is one of the frontier areas in hurricane research.
Research plan: We propose to run a 40-member Ens-DA and forecast cycle for hurricane Rita
2005. We will run WRF on a spatially extensive domain centered on the northern Caribbean Sea with
15-30 km resolution. A version of the Ens-DA scheme in the Data Assimilation Research Testbed
(http://www.image.ucar.edu/DAReS/DART) will be used and tested against a control run using the
WRF 3D-Var. We will assimilate the standard conventional operational observations, as well as
observations collected by the NOAA G-IV and WP-3D aircrafts.
In the first experiment, we will run Ens-DA with a single set of the physical parameterization
schemes. In other words, the ensemble only reflects the errors in the initial condition. To determine if
the Ens-DA improves the analyses relative to the 3D-Var, we will run forecasts initialized from the
Ens-DA’s ensemble mean analyses and compare with those initialized from the 3D-Var analysis.
Common hurricane verification metrics such as the track and intensity error will be evaluated against
the National Hurricane Center Best Track data.
In the second experiment, we will run Ens-DA with a multiple combinations of physics
parameterization schemes. In particular, we will combine several bulk microphysical schemes in
conjunction with a number of convective schemes available in WRF. We will then determine
whether the inclusion of the multi-physics ensembles in Ens-DA will further improve the analysis.
Expected outcome and impact: The end results are the knowledge of the feasibility of using Ens-
DA for hurricane initialization and probabilistic forecasts, and the knowledge of hurricane dynamics.
One major broader impact is the demonstration of an improved technique for operational hurricane
forecasts.
|
