Prediction, Assimilation & Risk Communication

Storm Surge Predictions

Storm surge is one of the most deadly and damaging hazards associated with landfalling hurricanes, and so NOAA and the research community have been working to improve storm surge predictions to help at-risk populations engage in timely protective action. In order to most effectively improve storm surge predictions, it is important to understand the potential predictive skill, as well as the most important sources of errors in current surge forecasts. To address this knowledge gap, MMM researchers have been using coupled hurricane - storm surge modeling to investigate the predictability of hurricane-induced storm surge across a range of lead times. The research uses the Advanced Circulation (ADCIRC) model to simulate the coastal inundation from storm surge produced by different realizations of landfalling tropical cyclones. As described in Fossell et al. (2017), this includes investigating the sensitivity of storm surge predictive skill to perturbations in four hurricane parameters — track, intensity, size, and translation speed — using location-specific and integrated metrics. Typical errors in hurricane forecasts are used to interpret the sensitivity results in terms of practical predictability of coastal inundation at different lead times.

In FY2018, MMM expanded this research beyond the two initial storms studied (Hurricanes Charley and Ike) to investigate two major storms that made landfall in the mainland U.S. during the 2017 Atlantic hurricane season (Hurricanes Harvey and Irma). Results across the storms indicate that location-specific surge is less predictable for small storms; given current average errors in hurricane forecasts, location-specific inundation from small storms like Charley and Harvey is at best predictable for only about 12 hours before landfall. Using an integrated inundation volume metric, however, suggests that there may be systematic relationships between storm size or intensity and total inundation (see figure).

Figure: Plot of normalized surge inundation volume vs. normalized maximum wind speed
Figure: Plot of normalized surge inundation volume (perturbed inundation volume / control inundation volume) vs. normalized maximum wind speed (perturbed v_max / control v_max - 1), for Hurricanes Ike, Charley, Irma, and Harvey. The experiments shown use the systematic storm perturbation approach and the storm-following inundation volume metric described in Fossell et al. (2017).

This research is partly funded by a research grant from the National Science Foundation (award #1331490) and involves collaboration with the National Oceanographic and Atmospheric Administration.





Assimilation of Radiance Data to Improve Forecasts

In a collaborative study (Wang et al., 2018) by scientists from NCAR/MMM, the Nanjing University of Information Science and Technology, and the China Meteorological Administration, the hourly geostationary AHI (Advanced Himawari Imager) radiance data from three channels measuring middle and upper troposphere moisture with a nominal resolution of 2-km were assimilated into convective-scale (3 km) WRF simulations. This substantially improved the forecast for a heavy rainfall event that occurred in the southwest region of Beijing (location A), where 24-accumulated rainfall exceeds 200 mm and the control experiment with the assimilation of only conventional observations significantly under-predicts the maximum rainfall center. In the southwest region of Hebei province (location B), both assimilation experiments show underprediction when comparing with the observed maximum rainfall exceeding 200 mm. However, the coverage of simulated rainfall exceeding 100 mm from the CONV+AHI experiment is improved and the rainfall pattern is closer to the observed one. Assimilating AHI data also corrected the northwestward location bias of another rainfall band in Hubei province (location C).

The observed 24-h accumulated rainfall (mm) over China from 1200 UTC 19 to 1200 UTC 20 July 2016
Figure: (a) The observed 24-h accumulated rainfall (mm) over China from 1200 UTC 19 to 1200 UTC 20 July 2016 and the corresponding WRF forecast rainfall at 3-km resolution from the experiment of assimilating (b) only conventional observations (CONV) and (c) conventional observations plus AHI (Advanced Himawari Imager) radiance data (CONV+AHI) using WRFDA-3DVAR.

Wang, Y., Z. Liu, S. Yang, J. Min, L. Chen, Y. Chen, and T. Zhang, 2018: Added value of assimilating Himawari-8 AHI water vapor radiances on analyses and forecasts for "7.19" severe storm over north China. J. Geophys. Res. Atmos., 123, https://doi.org/10.1002/2017JD027697





Cycled Data Assimilation used to Diagnose Systematic Model Errors

Improving our ability to predict the atmosphere and earth system, whether it is the details of tomorrow’s weather or climatic conditions in the coming decades, depends on continued improvements in the underlying numerical models used to simulate atmospheric evolution. Achieving this is nontrivial, because of the models’ complexity and computational cost, and because only limited observations of the atmosphere are available. One approach is through cycled data assimilation, which begins with a short model forecast (six-hour, in the work shown here) compared to observations valid at the same time. The predicted atmospheric state is updated to reflect the new observations, and this cycle is repeated indefinitely, with a new short forecast, followed by updating given observations at that subsequent time, etc.

The updates that are made to the predicted atmospheric state are called analysis increments. Systematic errors in the model forecasts are then manifested as nonzero increments in one or more of the model variables (for example, correcting for a warm or moist model bias at a given height) when averaged over many of these short-range forecasts. Such non-zero time-averaged increments result when the model continually drifts away from the corrections imposed by the observations in each assimilation cycle. This model drift can be seen in the model tendencies, and the total tendency for any model variable can then be decomposed into contributions from different elements of the numerical forecast model, such as the resolved (fluid) dynamics or additional physical processes ranging from microphysics (condensation, evaporation, and other aspects of phase changes of water) to unresolved turbulent fluxes to radiative heating and cooling. The decomposed tendencies help identify which elements of the forecast model may be the sources of systematic errors. As such, a cycled data assimilation system can serve as a useful platform to objectively evaluate and diagnose systematic model errors.

 A comparison of model biases of an updated cumulus scheme in a 15-km WRF forecast system covering the continental US
Figure: A comparison of model biases of an updated cumulus scheme (Tiedtke vs New Tiedtke) in a 15-km WRF forecast system covering the continental US. Model biases of the 6-h forecasts from a continuously cycling system using the Data Assimilation Research Testbed (DART) show improvements in the mid-tropospheric moisture bias using the New Tiedtke scheme. Biases are computed using available radiosonde measurements at 00 and 12 UTC over the Central US during a spring period. Decomposition of the model moisture tendencies shows that the cumulus scheme and its interactions with the microphysics scheme are likely driving the improvements.