Improve robustness of wind field estimation #174

thomas-fred · 2024-01-30T15:59:36Z

Deduplicate synthetic storm tracks using track_id and timestep
Deduplicate storm tracks prior to labelling with DatetimeIndex
Check storm tracks have consecutive and monotonically increasing timestep
Increase country-storm search buffer from 2 to 3 degrees
Filter storm tracks so we keep track sections from 'first arrival to last departure' rather than simply those intersecting a buffer
Limit interpolation to 6 hours between track timesteps. Otherwise return zero wind speeds for timestep.
Check wind field estimation parameters are physical
Refactor wind estimation so advective and rotational components are calculated more closely in code (eases re-use of evaluation grid)
Taper advective winds as function of absolute distance from storm eye -- advective field decays to nothing at radius 1,000km.
Improve wind field plotting
Update tests

The estimated wind field for each time step is a vector sum of two components, the advective field due to the translational motion of the storm and the rotational field. The rotational field intensity is a function of radius from storm eye. Previously, the advective field had no spatial dependence. This meant the entire spatial domain had some sort of background wind field (typically(!) a few m/s). In the event of a fast moving storm (say a decayed cyclone transitting the NA, this speed may be quite significant. For large domains like the USA, this could mean not insignificant background wind speeds many thousands of miles from the storm. I have decided that preventing this is worth the necessary assumptions and extra complexity. We now take the advective wind vector and multiply it by some reduction factor. This factor is the distance to storm eye, normalised by the radius to maximum winds, then tapered to 0 using a hyperbolic tangent function.

Previously, we used distance normalised by radius to maximum winds. Now, fade to 0 at 1,000km radius. Prevents a storm erroneously influencing the other side of a continent.

thomas-fred added 16 commits January 24, 2024 11:24

Increase track search radius

16d82a0

Fade advective winds as function of distance from eye

a9a4a92

Previously, we used distance normalised by radius to maximum winds. Now, fade to 0 at 1,000km radius. Prevents a storm erroneously influencing the other side of a continent.

Check estimate_wind_field input validity

ea1bb40

Limit track interpolation to ~6 hour gaps

f34772e

Refine plot style and discretisation

b0e5869

Handle failure of interpolate_track or estimate_wind_field

9b70565

Ensure filtered storm tracks are continuous in timestep

af80640

Permit stationary eye

b9abc4d

More descriptive logging message

fba50f4

Do not deduplicate IBTrACS

a57fe2b

Deduplicate prior to adding datetime index

8870c7c

Assert timesteps are consecutive natural numbers

9bad54b

Drop superfluous parameter

aac913c

Update tests to wind.py API changes

9151a9d

Update integration test output (minor value changes)

9d6a1d5

thomas-fred merged commit 62cbfe5 into main Jan 30, 2024
1 check passed

tomalrussell deleted the feature/taper_advective_winds branch January 31, 2024 10:49

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve robustness of wind field estimation #174

Improve robustness of wind field estimation #174

thomas-fred commented Jan 30, 2024 •

edited

Loading

Improve robustness of wind field estimation #174

Improve robustness of wind field estimation #174

Conversation

thomas-fred commented Jan 30, 2024 • edited Loading

thomas-fred commented Jan 30, 2024 •

edited

Loading