NATS 101 Lecture 26 Weather Forecasting 2

1
NATS 101Lecture 26Weather Forecasting 2

2
Review Key Concepts

• There are several types of forecasts
• Numerical Weather Prediction (NWP)
• Use computer models to forecast weather
• -Analysis Phase ?
• -Prediction Phase ?
• -Post-Processing Phase ?
• Humans modify computer forecasts

3
Suite of Official NWS Forecasts
CPC Predictions Page
4
3-Month SST Forecast Most recent

• SST forecasts for the El Nino region of tropical
  Pacific are a crucial component of seasonal and
  yearly forecasts.
• Forecasts of El Nino and La Nina show skill out
  to around 12 months.
• 1997-98 El Nino forecast was fairly accurate once
  El Nino was established

Strong La Nina
5
Winter 2007-2008 Outlooklatest prediction
6
Winter 2004-2005 Outlook(Issued 18 March 2004)
7
Winter 2004-2005 Outlook(Issued 18 March 2004)
8
Day 5 and 7 GFS Model
9
60 h ETA Forecast(Valid 0000 UTC 5 NOV 2001)

• NCEP model with finest resolution (12 km grid)
• ETA model gives the best precipitation forecasts

10
NCEP GFS Forecasts

• ATMO GFS Link
• NCEP global forecast 4 times per day
• Run on 50 km grid (approximately)
• GFS gives the best 2-10 day forecasts

11
NCEP GFS Forecasts

• ATMO NAM Link
• NCEP CONUS forecast 4 times per day
• Run on 12 km grid (approximately)
• NAM gives the best 24 h precip forecasts

12
Different Forecast Models
Ahrens 2nd Ed. Akin to Fig 9.1

• Different, but equally defensible models produce
  different forecast evolutions for the same event.
• Although details of the evolutions differ, the
  large-waves usually evolve very similarly out to
  2 days.

AVN-ETA-NGM Comparison
13
Forecast EvaluationAccuracy and Skill

• Accuracy measures the closeness of a forecast
  value to a verifying observation
• Accuracy can be measured by many metrics
• Skill compares the accuracy of a forecast against
  the accuracy of a competing forecast
• A forecast must beat simple competitors
• Persistence, Climatology, Random, etc.
• If forecasts consistently beat these
  competitors, then the forecasts are said to be
  skillful

14
Example of Accuracy Estimate
Ahrens 2nd Ed.
5 Day Forecast
Verification

• Absolute Error Forecast Value - Observed
  Value
• Error (Tucson) 5750 m-5780 m -30 m
  30 m
• Error (Newfoundland) 5280 m-5540 m
  -260 m 260 m
• Map average value is around 60 m, a sufficiently
  small error that the locations of the trough and
  ridge are accurately forecast

15
Example of Skill Estimate
Ahrens 2nd Ed.
5 Day Forecast
Verification

• Absolute Error (Tucson) 5750 m-5780 m
  -30 m 30 m
• Absolute Error (Climatology) 5690 m-5780 m
  -90 m 90 m
• The error for the model is less than the error
  for the climatology forecast, so the forecast is
  said to be skillful relative to climatology.

16
Current NWP Performance
Seasonal variation in skill for ETA rainfall
forecasts
Skill of NCEP models for rain
Aguado and Burt
24 h rainfall forecasts are skillful. Skill
decreases with rain amount. Skill varies with
season and year. Summer is most difficult season.
Aguado and Burt
17
How Humans Improve Forecasts

• Local geography in models is smoothed out.
• Model forecasts contain small, regional biases.
• Model surface temperatures must be adjusted, and
  local rainfall probabilities must be forecast
  based on experience and statistical models.
• Small-scale features, such as thunderstorms, must
  be inferred from long-time experience.
• If model forecast appears systematically off,
  human corrects it using current information.

18
Humans Improve Model Forecasts

• Forecasters perform better than automated model
  and statistical forecasts for 24 and 48 h.
• Human forecasters play an important role in the
  forecasting process, especially during severe
  weather situations that impact public safety.

Max Temp Accuracy
Aguado and Burt
Rainfall Skill
19
Current Skill

• 0-12 hrs Can track individual severe storms
• 12-48 hrs Can predict daily weather changes
  well, including regions threatened by severe
  weather.
• 3-5 days Can predict major winter storms,
  excessive heat and cold snaps. Rainfall forecasts
  are less accurate.
• 6-15 days Can predict average temp and rain over
  5 day period well, but daily changes are not
  forecast well.
• 30-90 days Some skill for average temp but not
  so much for rainfall over period. Forecasts use
  combination of model forecasts and statistical
  relationships (e.g. El Nino).
• 90-360 days Slight skill for SST anomalies.

20
Why NWP Forecasts Go Awry

• There are inherent flaws in all NWP models that
  limit the accuracy and skill of forecasts
• Computer models idealize the atmosphere
• Assumptions can be on target for some situations
  and way off target for others

21
Why NWP Forecasts Go Awry

• All analyses contain errors
• Regions with sparse or low quality observations
• - Oceans have poorer data than continents
• Instruments contain measurement error
• - A 20oC reading does not exactly equal 20oC
• Even a precise measurement at a point location
  might not accurately represent the big picture
• - Radiosonde ascent through isolated cumulus

22
Why NWP Forecasts Go Awry

• Insufficient resolution
• Weather features smaller than the grid point
  spacing do not exist in computer forecasts
• Interactions between the resolved larger scales
  and the excluded smaller scales are absent
• Inadequate representations of physical processes
  such as friction and heating
• Energy and moisture transfer at the earth's
  surface are not precisely known

23
Chaos Limits to Forecasting

• We now know that even if our models were perfect,
  it would still be impossible to predict precisely
  winter storms beyond 10-14 days
• There are countless, undetected small errors in
  our initial analyses of the atmosphere
• These small disturbances grow with time as the
  computer projects farther into the future
• Lorenz posed, Does the flap of a butterflys
  wings in Brazil set off a tornado in Texas?

24
Chaos Limits to Forecasting

• After a few days, these initial imperfections
  dominate forecasts, rendering it useless.
• Chaotic physical systems are characterized by
  unpredictable behavior due to their sensitivity
  to small changes in initial state.
• Evolutions of chaotic systems in nature might
  appear random, but they are bounded.
• Although bounded, they are unpredictable.

25
Chaos Kleenex Example

• Drop a Kleenex to the floor
• Drop a 2nd Kleenex, releasing it from the
  same spot
• Drop a 3rd Kleenex, releasing it from the
  same spot, etc.
• Repeat procedure1,000,000 times if you
  like, even try moving closer to the floor
• Does a Kleenex ever land in the same place as a
  prior drop?
• Kleenex exhibits chaotic behavior!

26
Atmospheric Predictability

• The atmosphere is like a falling Kleenex!
• The uncertainty in the initial conditions grow
  during the evolution of a weather forecast.
• So a point forecast made for a long time will
  ultimately be worthless, no better than a guess!
• There is a limited amount of predictability,
  but only for a short period of time.
• Loss of predictability is an attribute of
  nature. It is not an artifact of computer models.

27
A Chaotic System Ski Slope
Courtesy R. Houze, following Lorenz (1993)

• Many systems in nature are unpredictable
• Consider a simple ski slope with moguls

28
A Chaotic System Ski Slope

• Imagine 7 skis released at top of slope.
• All skis point in the same direction and have the
  same velocity, but they start from points
  separated by 10 cm along top of hill.
• Paths can be computed from Newtons 2nd Law and
  the relevant forces of gravity and friction.
• The results (on next page) show that the final
  positions of the skis are unpredictable.

29

• Positions at bottom of hill are much farther
  apart than at top of hill.
• Final positions of skis are very sensitive to
  their initial positions.
• If there is uncertainty in initial position, the
  final position is unpredictable.
• Example of Chaotic System
• The Atmosphere is Chaotic!

All ski tracks are closely bunched prior to 17 m
Ski tracks are widely spaced after 17 m
Lorenz 1993
30
A Smooth Ski Slope

• Now consider a smooth slope with no moguls.
• The skis would go downhill in a straight line.
• The final positions of the skis would always
  remain 70 cm apart, spaced at 10 cm intervals.
• Uncertainty in the final prediction, regardless
  of the forecast length, is no greater than the
  uncertainty in the initial positions of the skis.
  
• A smooth slope is not a chaotic system.

31
Ski Slope

• Although a chaotic system is ultimately
  unpredictable, it is somewhat predictable early.
• Note that the skis are closely spaced to 17 m.
  So the positions are fairly predictable at first.
• After 17 m, the paths diverge greatly and there
  is a loss of predictability.
• The skis have limited predictability.

32
Atmospheric Predictability

• The atmosphere is like the ski slope with moguls!
• The uncertainty in the initial conditions grow
  during the evolution of a weather forecast.
• So a pinpoint forecast made for a long time in
  the future is worthless, no better than a guess!
• There is a limited amount of predictability,
  but only for a short period of time.
• Loss of predictability is an attribute of
  nature. It is not an artifact of computer models.

33
Limits of Predictability

• What determines the limits of predictability for
  the atmosphere?
• Limits dependent on many factors such as
• Flow regime
• Geographic location
• Spatial scale of disturbance
• Weather element

34
Sensitivity to Initial Conditions
DAY 3 FORECAST POSITIVE PERTURB
DAY 3 FORECAST NEGATIVE PERTURB
VERIFYING ANALYSIS
DAY 3 FORECAST NOT PERTURBED
35
Sensitivity to Initial Conditions
DAY 3 FORECAST POSITIVE
DAY 3 FORECAST NEGATIVE
VERIFYING ANALYSIS
DAY 3 FORECAST UNPERTURBED
36
Summary Key Concepts

• NCEP issues forecasts out to a season.
• Human forecasters improve NWP forecasts.
• NWP forecast go awry for several reasons
• measurement and analysis errors
• insufficient model resolution
• incomplete understanding of physics
• chaotic behavior and predictability
• Chaos always limits forecast skill.

37
Assignment for Next Lecture

• Topic - Thunderstorms
• Reading - Ahrens pg 257-271
• Problems
• 10.1, 10.3, 10.4, 10.5, 10.6, 10.7, 10.16