Fog events occur at Melbourne Airport, Aus- tralia, approximately 12 times each year. Un- forecast events are costly to the aviation industry, cause disruption and are a safety risk. Thus, there is a need to improve operational fog forecasting. However, fog events are difficult to forecast due to the complexity of the physical processes and the impact of local geography and weather ele- ments. Bayesian networks (BNs) are a probabilistic rea- soning tool widely used for prediction, diagno- sis and risk assessment in a range of application domains. Several BNs for probabilistic weather prediction have been previously reported, but to date none have included an explicit forecast de- cision component and none have been used for operational weather forecasting. A Bayesian De- cision Network (Bayesian Objective Fog Fore- cast Information Network; BOFFIN) has been developed for fog forecasting at Melbourne Air- port based on 34 years of data (1972-2005). Pa- rameters were calibrated to ensure that the net- work had equivalent or better performance to prior operational forecast methods, which lead to its adoption as an operational decision sup- port tool. The operational use of the network by forecasters over an 8 year period (2006-2013) has been evaluated [1], showing significantly im- proved forecasting accuracy by the forecasters using the network, as compared with previous years. BOFFIN-Melbourne has been accepted by forecasters due to its skill, visualisation and explanation facilities, and because it offers fore- casters control over inputs where a predictor is considered unreliable... Paper: http://abnms.org/conferences/abnms2015/papers/ABNMS_2015_paper_16%20-%20A%20Tool%20for%20Visualising%20the%20output%20of%20a%20DBN%20for%20fog%20forecasting.pdf Tali Boneh, Xuhui Zhang, Ann Nicholson and Kevin Korb Seventh Annual Conference of the Australasian Bayesian Network Modelling Society conference (ABNMS2015) - Monash Caulfield, Melbourne, Australia 2015: http://abnms.org/conferences/abnms2015/

1. A Tool for Visualising the output of a
DBN for fog forecasting
Information Technology
Tali Boneh, Xuhui Zhang,
Ann Nicholson & Kevin Korb
Faculty of Information Technology
Monash University

2. Overview
• The problem: fog forecasting (at airports)
• Previous BN modelling:
– “static” Bayesian decision network (no explicit time)
– Operational decision support tool since 2006
• DBN for fog forecasting
• Fog DBN Visualisation tool

3. Fog Formation

4. Fog Forecasting for Airlines
• Terminal Aerodrome Forecast (TAF) issued every 6
hours; valid for 30h
3pm ---------------------------------| ~ 9am (Perth) /11am (Melbourne)
9pm ------------------------| ~ 9am
3am ---------------| ~ 9am
• When fog probability ≥ 30%, fog included in the TAF
• If fog forecasted, aircrafts must carry enough fuel to
– reach an alternative airport
– maintain a holding pattern above the airport
• When chance of fog >5% but <30%, forecast is “Code
Grey”

5. ARC Linkage
2002-2004
2005 2006- 2008 2010200920042003 2011 20152012 2013
Boneh’s PhD
2003-2008
ARC Linkage
2012-2014 (2015)
Boneh post-doc
Boneh @ BoM
2002 2014
1. ARC Linkage 2012-2014. Improving Meteorological Forecasting
with Knowledge Management Systems.
2. ARC Linkage 2012-2014.Temporal and spatial Bayesian network
modelling for improved fog forecasting
ARC $224K, BoM $148K
Boneh , T. (2010) Ontology and Bayesian Decision Networks for
Supporting the Meteorological Forecasting Process. PhD thesis,
Monash and Australian Bureau of Meteorology
collaboration 2003-2015
BOFFIN decision support tool for forecasting fog at
Melbourne airport

6. Phase 1: 2004 – 2008
(static Bayesian Decision Network)
Four priority airports were chosen for fog forecast improvement:
• Melbourne (~ 12-13 fogs)
– Networks for 3pm, 6/9pm,
midnight
• Sydney (~ of 4 to 5 fogs per year,
largest traffic volumes),
– Networks for midnight, 2am and
3am
• Canberra (~ 42 fogs)
– Networks for 6pm, midnight and
3am
• Perth (~ 12 fogs per year, large
distances to the nearest alternate airports
– A network for 3pm
Prediction: Fog Yes or No?
No prediction of onset and clearance
Month
January
February
March
April
May
June
July
August
September
October
November
December
8.46
7.71
8.46
8.18
8.46
8.18
8.70
8.57
8.18
8.46
8.18
8.46
LengthOfNight
Nov to Jan
Feb and Oct
March and Sept
Apr and Aug
May to July
25.1
16.2
16.6
16.8
25.3
RainNoRain
0 to 4.5
>= 4.5
90.6
9.41
Fog
fog
nofog
3.30
96.7
U
LapseRate9pmCont
< 2.05
2.05 to 2.75
2.75 to 3.25
>= 3.25
26.2
18.1
17.7
37.9
2.74 ± 0.75
Moisture
Vfav
fav
unfav
26.4
19.2
54.4
SternParkyn
0 to 1
1 to 2
2 to 5
5 to 10
10 to 15
15 to 30
30 to 100
46.8
15.1
17.5
9.79
4.12
4.40
2.28
4.8 ± 11
Decision
saynofog
CodeGreyLessThan5
CodeGrey5
CodeGrey10
CodeGrey20
sayfog
28.7044
28.9378
28.0293
26.3915
24.6700
22.6587
Gradient
Vfav
fav
unfav
31.2
19.3
49.4

7. “Static” semi-causal BN for fog forecasting
Decision
saynofog
CodeGreyLessThan5
CodeGrey5
CodeGrey10
CodeGrey20
sayfog
28.7044
28.9378
28.0293
26.3915
24.6700
22.6587
Month
January
February
March
April
May
June
July
August
September
October
November
December
8.46
7.71
8.46
8.18
8.46
8.18
8.70
8.57
8.18
8.46
8.18
8.46
LengthOfNight
Nov to Jan
Feb and Oct
March and Sept
Apr and Aug
May to July
25.1
16.2
16.6
16.8
25.3
RainNoRain
0 to 4.5
>= 4.5
90.6
9.41
Fog
fog
nofog
3.30
96.7
U
LapseRate9pmCont
< 2.05
2.05 to 2.75
2.75 to 3.25
>= 3.25
26.2
18.1
17.7
37.9
2.74 ± 0.75
Moisture
Vfav
fav
unfav
26.4
19.2
54.4
SternParkyn
0 to 1
1 to 2
2 to 5
5 to 10
10 to 15
15 to 30
30 to 100
46.8
15.1
17.5
9.79
4.12
4.40
2.28
4.8 ± 11
Gradient
Vfav
fav
unfav
31.2
19.3
49.4
Background
Decision
Guidance
Predictors
Target

8. BOFFIN: deployed 2006
• Decision support tool in operational use for fog
forecasting at Melbourne airport
– BN as back-end reasoning engine
– Integrated with data stream of current weather
observations
– Allows forecasters to adjust data entered as evidence
(“what-if” functionality)
• Improved forecasting 2006-2014 (fewer missed fogs
without increase in false alarms)
Boneh et al. (To Appear). Weather and Forecasting. Fog Forecasting
for Melbourne Airport using a Bayesian Decision Network.
• Limitation: no temporal aspect to prediction, i.e.
time of onset and clearance

9. Dynamic Bayesian networks
• For explicit reasoning over time, a copy of each process
variable X at each time T: X0, X1, … XT-1, XT, XT+1, …
• Typically:
– Same structure within each time slice (intra-slice arcs)
– Stationary time series (i.e. parameters not time dependent), so
summarised in two time slice DBN (with sliding window), used
with “roll-up, roll-out” approximation

10. Weather
Fog
Weather
Fog
Weather
Fog
Clearance
Onset
times
Clearance
Onset
times
Weather Weather
Forecast time
Now
Forecast time
Future
Forecast time
Future
NWP
Predictions
Current
Obs
NWP
Predictions
Current
Obs
NWP
Predictions
DBN Framework for Fog

11. Weather
Fog
Weather
Fog
Weather
Fog
Clearance
Onset
times
Clearance
Onset
times
Weather Weather
Forecast time
Now
Forecast time
Future
Forecast time
Future
NWP
Predictions
Current
Obs
NWP
Predictions
Current
Obs
NWP
Predictions
DBN Prototype 1Add observations
directly
Non-causal (NB)

12. 12
Melbourne Fog DBN: Prototype 1

13. 14
5 Weather variables (discretised): children of
both “length of night” and the Fog node

14. 15
Melbourne Fog DBN: Prototype 1

15. 16
Change nodes:
If Fog, Clearance? Time of clearance?
If NoFog, Onset? Time of Onset?
If NoFog, then Onset, will it clear?

16. Melbourne Fog DBN: Prototype 1
17
Visualisation and understandability: 8 time slices
(Netica demo)

17. ForecastTime Info.
Nodes midN 3am 6am
F_F_TmidN ~{fog} nofog1 nofog1
delta_F1_TmidN * onset onset
delta_F2_TmidN * nochange nochange
Onset_TmidN * 0.25 0.25
Clear_if_Clearance_TmidN * * *
Clear_if_Onset_TmidN * * *
Wvis_TmidN 1 1 1
F_F_T3am * fog fog
delta_F1_T3am * * clearance
delta_F2_T3am * * *
Onset_T3am * * *
Clear_if_Clearance_T3am * * 1
Clear_if_Onset_T3am * * *
Wvis_T3am * 1 1
F_F_T6am * * ~{fog}
delta_F1_T6am * * *
delta_F2_T6am * * *
Onset_T6am * * *
Clear_if_Clearance_T6am * * *
Clear_if_Onset_T6am * * *
Wvis_T6am * * 4
First attempt at presenting DBN outputs
A fog case (fog 00:15 - 4:00)
midN
3am
6am
net slice

18. First attempt at presenting DBN outputs
A fog case (fog 00:15 - 4:00)
probabilities of ‘fog now’ at different forecast times
forecast time 9pm
0.00100943
0.0013859
0.00210729
0.00293861
0.00580142
0.0107448
0.0203179
0.0339942
0
0.00858423
0.0214569
0.0355907
0.0516443
Probabilities
Start interval Time
forecast time
3am
0.00100943
0.0013859
0.00210729
0.00293861
0.00580142
0.0107448
0.0203179
0.0339942
1
0.977578
0.903714
Probabilities
Start interval Time
forecast time midN
0.00100943
0.0013859
0.00210729
0.00293861
0.00580142
0.0107448
0.0203179
0.0339942
0
0.0393763
0.0479343
0.0538041
Probabilities
Start interval Time 0.00100943
0.0013859
0.00210729
0.00293861
0.00580142
0.0107448
0.0203179
0.0339942
0
0.0799313
Probabilities
Start interval Time
Prior 6am
Forecast Time
forecast time 6am
forecast time 6pm
0.00100943
0.0013859
0.00210729
0.00293861
0.00580142
0.0107448
0.0203179
0.0339942
0
0.000694723
0.00468568
0.0129131
0.0275323
0.0488081
Probabilities
Start interval Time
• Our forecasters had no idea what the DBN was doing/saying
• And we weren’t sure either!

19. Fog DBN Visualisation Tool
• Developed using D3, a JavaScript library for
manipulating documents based on data within a web
browser, uses a combination of HTML, SVG, and CSS
• Extends Matthew Weber’s “block” template
• Developed in collaboration with visualisation
research group at Monash

20. Selected Cases for Demo
22
# Actual DBN date onset clearance Comment and
forecast at 3pm
1 fog fog
hit
28/06/2011 19:00 2:30 Always looked like
fog (TAF)
2 no fog no fog
correct
4/07/2003 Always looked like
no fog (TAF)
3 fog no fog
miss
3/07/2008 7:21 8:00 Looked like no fog
until 6am (CG)
4 no fog fog
false alarm
21/05/2009 Fog on previous day
cleared at 9:47am
(TAF)

21. X-axis: time over night
Y-axis: P(X=x|Observations to date)
30% chance of fog
Prior Select
X=x
To
display
Case “fog hit” (fog event successfully forecast):
sequence of observations for weather variables and fog, from
midday to following 9am
Forecast time – here midday
Visualisation
of forecast
time

22. fogTonight: hovering over point brings up probability
Note: Fog prior not selected but is always shown
“faded”

23. Displaying more info: e.g. probability of fogOnset

24. Displaying more info: e.g. probability of fogOnset

25. Overlay showing not only probability for fogOnset in next 3
hrs, but distribution of times for onset and clearance

26. Overlay showing not only probability for fogOnset in next 3
hrs, but distribution of times for onset and clearance

27. Probability of “fogNow”, option for overlay with
distribution for time of fog clearance

28. Demo

29. No fog case, correctly predicted

32. No fog case, false alarm

33. Showing probabilities for single value, at different
forecast times

34. No fog case, false alarm

35. No fog case, false alarm

36. Where are we at?
• Bur. of Met collaborators ‘like’ the DBN visualisation tool
– Understanding of the DBN model
– Assisting with ongoing development
• No usability testing as such
• Currently doing lots of experimental work on DBN
– Selection of weather variables as predictors
– Discretisation of weather variables, onset and clearance times
• Specific research project outcomes:
– prototype DBN, papers
• Future: embed DBN in support tool for operational use
(1-2 years?)