Chapter 12 - Operational NWP.pdf

Chapter 12
Operational NWP

Operational models are run by:
• Large weather services – e.g., NCEP, ECMWF,
UKMO, DWD…
• Universities – for educational purposes
• Insurance companies
• Consulting companies
– general high-res weather providers
– Specialized products: air-quality, aviation, road weather,
agriculture, maritime
• Small countries
• Defense-related organizations – Navy, Army, Air
Force

A simple operational modeling system
Specialized
secondary models

Operational weather vs
operational climate prediction
• There is no well-defined forecast-length
threshold that distinguishes the two terms
– e.g., interseasonal prediction: is it
weather or climate?
• Decadal climate predictions will soon be
initialized like weather forecasts, so there is
no distinction in that sense.

Model reliability – a distinction
between research and operational apps
• Research apps – if the model crashes,
you just fix the problem and rerun
• Operational apps – a crash in a
sequential initialization process is
difficult to gracefully recover from,
and may require a cold start.

Causes for model crashes
and how to avoid them
• Computer hardware problems – hard to
anticipate.
• CFL errors – Using a very short Δt to guarantee
that no violations will ever occur wastes
computer time (that could be used for higher
resol). But the Δt for an operational model
should be somewhat smaller than for a
research model.

• Transmission of observations may be
interrupted – but, we can initialize from the
background only, if needed.
• LBCs for LAMs – If the global model does not
complete on time, or at all, the cycle for the
LAM is compromised.
• Severe LBC noise with LAMs can cause a crash.
• Physical-process parameterizations may not
work for an unusual combination of inputs.

Other considerations related to LAMs
• The initialization is delayed for a few hours
because it is necessary to wait for the global
model to finish its forecast (LBCs).

Factors that control
model execution speed
• Must wait for most observations to arrive – model “cutoff”
times can be 60-90 minutes.
• For LAMs, must wait for the LBCs
• The time step – requirement depends on numerical
approach, and how conservatively the time step is chosen if
there is a constraint.
• The number of grid points
• The speed of the processors, the number of processors,
and how well the execution time scales with the number of
processors.
• The sophistications of the physics parameterizations
• Make output available as the model is running.
• Execution speed versus user friendliness
• Large I/O loads can slow execution speed.

Types of post-processing
of the met forecasts
• Bias correction (Chap 13)
• Statistical diagnosis of variables that are
not forecast well – fog, freezing rain,
turbulence intensity
• Secondary models (Chap 14)

Miscellaneous topics
• Managing model development – requires
multiple computer systems
– Primary forecast system
– System for parallel runs to test upgrades
– R&D system
• Real-time verification
• The relative roles of models and forecasters

Properties of user-friendly
operational displays