Improved Low Visibility forecast

1
Improved Low Visibility forecast At Amsterdam
Airport Schiphol
Nico Maat Daan Vogelezang
May 2008
Photo Peter de Vries
2
KDC Knowledge Development Centre (foundation)

• Members of Aviation Sector
• KLM airlines, LVNL (ATC), Amsterdam Airport
  Schiphol (AAS)
• Consulted Parties NLR, KNMI, Boeing, TU Delft,
  To70, ADSE, Maastricht UAC,
• Mission
• To integrate knowledge and to develop solutions
  which enable to ensure and expand the mainport
  position of Schiphol airport in Europe.

3
The KDC-LVP Project (part I)

• Reduced Visibility gt Reduced Runway Capacity
• Larger in trail separations
• Less usable runway combinations
• Low visibility forecast gt flow restrictions
  enforced
• Incorrect visibility forecast gt
• Unnecessary flow restrictions (false alarm)
• Too late flow restrictions (miss)
• Project Goal
• Increase accuracy and reliability of low
  visibility (LVP) forecasts

4
LVP Conditions Flow Restrictions
5
Aeronautical Forecast cascade
SKV Schiphol Probability Forecast
6
Improvements TAFG

• RVR in addition to MOR
• More -closer- upstream sites advection
  predictors
• Joint probabilities for visibility and ceiling
• Reduced residual variance in dependent data

7
RVR as a function of MOR and BG luminance
8
Example MOR/BGL/RVR timeseries
9
Improvements TAFG MOR-gt RVR

10
Old vs. New BZO forecastBZO Limited Visibility
Operations

• old BZO prob. forecast
• combination of independent CLB threshold prob.
  and MOR threshold prob. gt take largest (full
  dependence)
• MOR-gtRVR translation using table and guess BGLum
• new BZO forecast
• Joint probs of MOR and CLB are calculated P (MOR
  lt X m or CLB lt Y ft)
• deterministic fc for BG Luminance
• prob.MOR threshold translated/interpolated to
  prob. RVR

11
Transformation
12
Verification of probabilities

• 3 years 200407-200704
• old independent data (TAFG SKV)
• new dependent data (TAFG)
• short term 8/day issue 4,7,10 (N24816)
• long term 4/day 9,12,,24 (N28952)

13
P lt LVP-A (short)
RVR lt 1500 m or ceiling lt 300 ft
14
P lt LVP-A (long)
15
From Probability to Decision

• How would the (my) decision have worked out in
  the past?
• What does a (wrong) forecast cost ?
• Several Options
• choose largest category
• choose category at a fixed P
• which P ?
• (choose worst if Prob(worst)gt X
• which X?)

16
Choice fixed percentage P25
200305 - 200704
current TAFG 02 4
Observation
17
verification current/new
p50
p25
note MG-MG is not considered in these figures
18
Users expense/cost Matrix
Forecast BZO category
Miss
Observed BZO category
False Alarm
gtFalse alarm and miss are equally costly
19
total historical expense
Forecast BZO category
Observed BZO category

x
Observation
sum of cells 781 total (historical) expense
gtgt
20
Expense for TAFG 02z 4

• Absolute cost much lower for new cf. current

21
Minimal expense gt Optimal P
blue largest category fa/mi1
22
Minimal expense gt Optimal P
blue largest category fa/mi1 green fa/mi1,
min. 25-70
23
Minimal expense gt Optimal P
blue largest category fa/mi1 green fa/mi1,
min. 25-70 red fa/mi10, min. 60-80
24
Minimal expense gt Optimal P
blue largest category fa/mi1 green fa/mi1,
min. 25-70 red fa/mi10, min. 60-80 purple
fa/mi0.1, min. 15-20
25
conclusions

• One probability forecast fits all users
• User may skip irrelevant categories
• User defines the cost of a wrong category
• optimize P is (in general) better than chosing
  largest category

outlook

• perform independent verification
• use real cost values