|
|
|
Predicting the Perfect Storm: Feature Identification and State Estimation for Nonlinear Systems
|
|
|
| We propose to develop new methods for representing and propagating statistical distributions that will encode multiple competing
hypotheses about weather system state. We allow the distributions representing states to undergo nonlinear evolution as time unfolds,
using known data to learn and track configurations of weather features. Furthermore, we propose to provide a conditional forecasting
capability, enabled by operating not just at the conventional level of grid cells but using interpretable features like fronts. By
conditioning on certain outcomes of particular concern (e.g., hurricane landing points), state trajectories likely to produce these
outcomes can be derived, and their relative likelihood computed to aid risk assessment. This conditioning capability is far more
efficient than current ensemble methods, especially in evaluating low-probability (but high-cost) outcomes. We will show how to link
the dense, gridded data NASA gathers into dynamic feature models comprehensible to humans, a capability NASA does not have. This
line of work could revolutionize atmospheric and oceanic forecasting capabilities by addressing directly the predictability of distinct
flow features, such as hurricanes, fronts and eddies, versus the approach of whole-field forecasting currently used in numerical
weather prediction. We offer a conditional forecast technology far more controllable and efficient than current ensemble-based
methods, and present plans to test it by identifying a low-probability formation that would be overlooked by conventional prediction
methods. |
|
|
