Analog forecasting of ceiling and visibility using fuzzy logic and data mining Bjarne Hansen Meteorological Research Branch Meteorological Service of Canada

Analog forecasting of ceiling and visibilityusing fuzzy logic and data mining

Bjarne Hansen

Meteorological Research Branch

Meteorological Service of Canada

Dorval, Quebec

Eastern Canada Aviation Weather Workshop

Montréal, Quebec, 16-18 September 2003

Introduction

Fuzzy Logic

Case-Based Reasoning

Weather Prediction

Ceiling and Visibility Prediction

Prediction System: WIND

Results

Conclusion

Future

Outline

By building expert systems that combine forecaster expertise,AI, large amounts of data (climatological and current),and currently available computing power, we can increase forecast quality and increase forecasting efficiency.

* Presentation at: http://iweb.cmc.ec.gc.ca/~armabha/metai

*

Basic computer science

Basic meteorology

Use of fuzzy logic has

increased exponentially

over the past 30 years,

based on the number of uses of

the word “fuzzy” in titles of

articles in engineering and

mathematical journals. 1

In meteorological systems,

use of fuzzy logic began

about ten years ago. 2

Fuzzy Logic

1. Lofti Zadeh, 2001: Statistics on the impact of fuzzy logic, http://www.cs.berkeley.edu/~zadeh/stimfl.html

2. Applications of fuzzy logic for nowcasting, http://chebucto.ca/Science/AIMET/applications

Fuzzy Logic Definition

1. Free On-line Dictionary of Computing, http://foldoc.doc.ic.ac.uk/foldoc

“Fuzzy logic a superset of Boolean logic dealing with the concept of

partial truth – truth values between ‘completely true’ and ‘completely false’.

It was introduced by Dr. Lotfi Zadeh of UCB in the 1960’s as

a means to model the uncertainty of natural language.” 1

0.00

0.25

0.50

0.75

1.00

-20 -10 0 10 20

difference (°C)

m

very

slightly

quite

Fuzzy set to describe

the degree to which

two numbers are

similar, for example,

degree of similarity

of temperatures.Simila

r

Not similar

Non-fuzzy (classical) setloses information aboutdegree of similarity.

1. Munakata, T. and Jani, Y., 1994: Fuzzy Systems: An Overview, Communications of the ACM, Vol. 37, No. 3, pp.69-76.2. Hansen et al. 1999, http://chebucto.ca/Science/AIMET/fuzzy_environment

Fuzzy logic is used in expert systems in many domains:

transportation, automobiles, consumer electronics, robotics,

pattern recognition, classification, telecommunications,

agriculture, medicine, management, and education. 1

Fuzzy logic often models continuous, real-world systems. There are

hundreds of fuzzy logic based systems that deal with environmental data.

agriculture, climatology, ecology, fisheries, geography, geology,

hydrology, meteorology, mining, natural resources, oceanography,

petroleum industry, risk analysis, and seismology. 2

Fuzzy Logic Applications

Meteorological view: CBR = analog forecasting

AI view: CBR = retrieval + analogy + adaptation + learning 1

CBR is a way to avoid the “knowledge acquisition problem.”

CBR is very effective in situations “where the acquisition

of the case-base and the determination of the features is

straightforward compared with the task of developing the

reasoning mechanism.” 2

CBR and analog forecasting recommended when models are

inadequate, e.g., ceiling and visibility, which are strongly

determined by local effects below scale of current computer models.

Case-Based Reasoning

1. Leake, D. B., 1996: CBR in context. The present and future; in Leake, D. B. (editor), Case-Based Reasoning: Experiences, Lessons & Future Directions, American Association for Artificial Intelligence, Menlo Park California, USA, 3-30.2. Cunningham, P., and Bonzano, A., 1999: Knowledge engineering issues in developing a case-based reasoning application, Knowledge-Based Systems, 12, 371-379.

k-Nearest Neighbor(s) Technique: k-nn

Definition: “For a particular point in question, in a population

of points, the k nearest points.” 1

Intuition: The closer the neighbors, the more useful they are for prediction.

“It is reasonable to assume that observations which are close together

(according to some appropriate metric) will have the same classification.

Furthermore, it is also reasonable to say that one might wish to weight the

evidence of a neighbor close to an unclassified observation more heavily than

the weight of another neighbor which is at a greater distance from the

unclassified observation.” 2

k-nn is a basic CBR method. Commonly used to explain an observation

when there is no other more effective method. 2

1. Dudani, S. A., 1976: The distance-weighted k-nearest neighbor rule, IEEE Transactions on Systems, Man, and Cybernetics, Volume SMC-6, Number 4, April 1976, 325-327.2. Aha, D. W. (1998) Feature weighting for lazy learning algorithms. In Liu, H. and Motoda, H. (Eds.), Feature Extraction, Construction, and Selection: A Data Mining Perspective. Norwell MA: Kluwer,

Fuzzy k-Nearest Neighbor(s) Technique: fuzzy k-nn

Definition: “Nearest neighbor technique in which the

basic measurement technique is fuzzy. 1

Two improvements to k-nn technique by using fuzzy k-nn approach: 1

“Improve performance of retrieval in terms of accuracy because of

avoidance of unrealistic absolute classification.”

“Increase the interpretability of results of retrieval because the

overall degree of membership of a case in a class that provides a

level of assurance to accompany the classification.”

1. Keller, J. M., Gray, M. R., and Givens Jr., J. A., 1985: A fuzzy k-nearest neighbor algorithm, IEEE Transactions on Systems, Man, and Cybernetics, Vol. 15, No. 4, 258-263.

Two basic methods to predict weather: 1

• Dynamical - based upon equations of the atmosphere, uses finite element

techniques, and is commonly referred to as computer modeling.

• Empirical - based upon the occurrence of analogs, or similar weather situations.

1. Lorenz, E. N., 1969: Three approaches to atmospheric predictability, Bulletin of the American Meteorological Society, 50, 345-349.

Weather Prediction

In practice, hybrid methods used: Models + Observations

Statistical methods infer estimated expected distributions under specified conditions.

Theoretical distributions are fit to sparse data, e.g. normal distributions, MLR.

Resampling methods are a recently feasible option, thanks to advances in computer

speed and storage, when data sets are large, and when condition specification is

deferred to the last minute (run time, time-zero), e.g., k-nn, WIND.

Statistical

Analog / Resampling

2. Rudner, Lawrence M. & Shafer, Mary Morello, 1992: Resampling a marriage of computers and statistics. Practical Assessment, Research & Evaluation, 3(5). http://EdResearch.org/pare/getvn.asp?v=3&n=5

Ceiling and Visibility PredictionCeiling height and visibility prediction demands precision:

Ceiling height, when low, accurate to within 100 feet.

Visibility, when low, accurate to within 1/4 mile.

Time of change of flying category should be

accurate to within one hour.

1. RAP/NCAR, Ceiling and visibility, Background, http://www.rap.ucar.edu/asr2002/j-c_v/j-ceiling-visibiltiy.htm

Safety concern

“Adverse ceiling and visibility conditions can produce major

negative impacts on aviation - as a contributing factor in

over 35% of all weather-related accidents in the U.S. civil aviation

sector and as a major cause of flight delays nationwide.” 1

Motivation for ceiling and visibility prediction research

Economics and Efficiency

Every 1% increase in TAF accuracy would result in $1M per year of value to

the air traffic system in Canada – estimating conservatively, and assuming

increase relative to recently measured levels of TAF accuracy. 1

The commonest cause for TAFs needing to be amended is the occurrence

of unforecast categories of cloud ceiling and visibility. 2

The National Weather Service (NWS) estimates that a 30 minute lead-time

for identifying cloud ceiling or visibility events could reduce the number of

weather-related delays by 20 to 35 percent and that this could save between

$500 million to $875 million annually. 3

“The economic benefit of a uniform, hypothetical increase in TAF accuracy

of 1% is approximately $1.2 million [Australian] per year for Qantas Intl.

flights into Sydney.” 4

1. Assessment of Aerodrome Forecast (TAF) Accuracy Improvement, NAV CANADA, May 2002, pg. 22.2. Henry Stanski, 1999: Personal communication.3. Jim Valdez, NWS Reinventing Goals for 2000, http://govinfo.library.unt.edu/npr/library/announc/npr5.htm4. Leigh, R. J., 1995: Economic benefits of Terminal Aerodrome Forecasts (TAFs) for Sydney Airport, Australia, Meteorological Applications, 2, 239-247.

Motivation for AI-based ceiling and visibility prediction research

Scientific and Engineering

Ceiling and visibility are “not resolvable” with current computer models

(aka NWP, numerical weather prediction models).

“Unfortunately, cloud cover is the most difficult of meteorological variables for

numerical models to predict. [MOS] output for predictions of ceiling and visibility

is heavily dependent on the most recent station observations rather than the

output of the numerical model. Consequently, the quality of ceiling and visibility

forecasts has not increased as it has for other forecast variables. For 3- and 6-

hour forecasts, several studies have shown that local forecasters could not do

better and often did worse than persistence. MOS forecasts were not clearly

better than those of the local forecaster for time frames of 9 hours or less.” 1

Persistence forecasting is a difficult technique to beat for very short-range

forecasting. 2 [Because of high ratio of VFR : IFR]

1. The COMET Outreach Program, http://www.comet.ucar.edu/outreach/9915808.htm

2. Dallavalle, J. P., and Dagostaro, V. J., 1995: The accuracy of ceiling and visibility forecasts produced by the National Weather Service, Preprints of the 6th Conference on Aviation Weather Systems, American Meteorological Society, 213-218

Prediction System: WIND

WIND: “Weather Is Not Discrete”

Consists of three parts: Data – weather observations and model-based guidance. Fuzzy similarity-measuring algorithm – small C program. Prediction composition – fairly trivial, predictions are based on selected percentiles of cumulative summaries of k nearest neighbors, k-nn.

Data Past airport weather observations, 32 years of hourly observations. Recent and current observations. NWP-based guidance.

Data: Past and current observations

Category

temporal

cloud ceilingand visibility

wind

precipitation

spread andtemperature

pressure

Attribute

date

hour

cloud amount(s)cloud ceiling heightvisibility

wind directionwind speed

precipitation typeprecipitation intensity

dew point temperaturedry bulb temperature

pressure trend

Units

Julian date of year (wraps around)

hours offset from sunrise/sunset

tenths of cloud cover (for each layer)height in metres of 6/10ths cloud coverhorizontal visibility in metres

degrees from true northknots

nil, rain, snow, etc.nil, light, moderate, heavy

degrees Celsiusdegrees Celsius

kiloPascal × hour -1

Data: Past and current observations

E.g., over 300,000 consecutive hourly obs for Halifax Airport, quality-controlled.

YY/MM/DD/HH Ceiling Vis Wind Wind Dry Dew MSL Station Cloud

Directn Speed Bulb Point Press Press Amount

30's m km 10's deg km/hr deg C deg C kPa kPa tenths Weather

64/ 1/ 2/ 0 15 24.1 14 16 -4.4 -5.6 101.07 99.31 10

64/ 1/ 2/ 1 13 6.1 14 26 -2.2 -2.8 100.72 98.96 10 ZR-

64/ 1/ 2/ 2 2 8.0 11 26 -1.1 -2.2 100.39 98.66 10 ZR-F

64/ 1/ 2/ 3 2 6.4 11 24 0.0 -0.6 100.09 98.36 10 ZR-F

64/ 1/ 2/ 4 2 4.8 11 32 1.1 0.6 99.63 97.90 10 R-F

64/ 1/ 2/ 5 2 3.2 14 48 2.8 2.2 99.20 97.50 10 R-F

64/ 1/ 2/ 6 3 1.2 16 40 3.9 3.9 98.92 97.22 10 R-F

64/ 1/ 2/ 7 2 2.0 20 40 4.4 4.4 98.78 97.08 10 F

64/ 1/ 2/ 8 2 4.8 20 35 3.9 3.3 98.70 97.01 10 F

64/ 1/ 2/ 9 4 4.0 20 29 3.3 2.8 98.65 96.96 10 R-F

64/ 1/ 2/10 6 8.0 20 35 2.8 2.2 98.60 96.91 10 F

64/ 1/ 2/11 8 8.0 20 32 2.8 2.2 98.45 96.77 10 F

64/ 1/ 2/12 9 9.7 23 29 2.2 1.7 98.43 96.75 10 F

64/ 1/ 2/13 9 11.3 23 32 1.7 1.1 98.37 96.69 10

...

Data: Computer model based guidance

a(t0)

b(t0)

a(t0-p)

b(t0-p)

guidance

b(t0+p)

Timezero

Recentpast Future

...

... ...

... ...... ... ...

Present Case

Past Cases

TraversingCase BaseSimilarity measurement

Prediction System – Data Structure and Case Retrieval

Compose present case: recent obs + NWP

Collect most similar past cases

m c (x 1 - x 2 )

0.00

0.25

0.50

0.75

1.00

-c 0 c

x 1 - x 2

very

slightlynear

Fuzzy similarity-measuring function

Three types of fuzzy operations designed to measure

degree of similarity between three types of attributes.

1. Continuous. (e.g., wind direction, temperature, etc.)

Design tight fit for critical elements, such as wind direction, relatively loose fit for others, such as temperature. An expert forecaster suggests values that

correspond to varying degrees of similarity.

Attribute slightly near near very near

date of the year 60 days 30 days 10 days

hour of the day 2 hours 1 hours 0.5 hours

wind direction 40 degrees 20 degrees 10 degrees

dew point temperature 4 degrees 2 degrees 1 degree

dry bulb temperature 8 degrees 4 degrees 2 degree

pressure trend 0.20 kPa hr -1 0.10 kPa hr -1 0.05 kPa hr -1

Expertly configured similarity-measuring function

Expert specifies thresholds for various degrees of near

0 8 16 24 320

8

16

24

32

m(x1,x2)

x1

x2

m(x1,x2)

0.75-1

0.5-0.75

0.25-0.5

0-0.25


2. Magnitude. (e.g., wind speed)

FuzzyDecisionSurface

Calm to lights wind speedsrequire special interpretation.0

04 8

4

8

Nil 1.00Drizzle 0.02 1.00

Showers 0.03 0.50 1.00Rain 0.01 0.50 0.75 1.00

… … … … … …Nil Drizzle Showers Rain …


3. Nominal. (e.g., precipitation)

Fuzzy Relationships

Different types of weather havedifferent perceiveddegrees of similarity.

Algorithm: Collect Most Similar Analogs, Make Prediction

.

. . .

Analogensemble

Climate archive

Archive search is like contracting hyperellipsoid centered on present case.

Axes measure differences weather elements between compared cases.

“Distances” determined by fuzzy similarity-measuring functions, expertly tuned, all applied together simultaneously.

C&V evolution

Forecast ceiling and visibilitybased on outcomes ofmost similar analogs.

Spread in analogs helps toinform about appropriateforecast confidence.

Prediction

WIND makes 11 series of deterministic forecasts based on

percentiles of C&V in analogs (0, 10, 20, ..., 100): 0%ile is the

lowest C&V, 50%ile is the median, 100%ile is the highest.

Using MSC / Nav Canada performance measures, experiments

showed that the series in the 20 to 40 range verified fairly well.

Be aware of systematic tradeoffs between Frequency of Hits,

False Alarm Ratio, and Probability of Detection, e.g.,

IFR POD and FAR

VFR POD and FAR

Prediction

Forecast: ceiling and visibility based on 30%ile of analogs

Prediction

Probabilistic forecast: 10 %ile to 50%ile cig. and vis. from analogs

Results

Forecasts are competitive with

persistence in 0-to-6 hour range,

and better than persistence in

the 0-to-24 hour range

based on FOH, FAR, and POD

of alternate and VFR forecasts. WIND runs in real-time for climatologically different sites.Data-mining/forecast processtakes about one second.

First impressions and

forecaster feedback:

Probabilistic forecasts of cig. & vis. informative, high “glance value”.

A “heads-up” message about the current “forecasting issue”

would be helpful, e.g., if wind > 8 knots and temp > -40oC

then Chance(ice fog) = Low.

Forecaster Feedback

1. WIND forecast blizzard conditions to improve to VFR after

one hour.

Analog ensemble used to base predictions on was too large,

as blizzards are a relatively rare event. Made a few changes to

the code and then WIND forecast blizzard conditions more

intelligently.

2. WIND often provides very good timing of significant category

changes.

Owe some credit to model guidance in many cases as, if wind

shifts and precipitation are well forecast by the model, WIND

benefits directly, and forecasts ceiling and visibility accordingly.

Forecaster Feedback

3. WIND provides reasonable values for the 6-to-24 hour period which could help in writing TAFs. Forecasting ceiling and visibility in this time period is presently difficult for forecasters because nowcasting techniques, such as persistence and extrapolation, are unreliable.

4. WIND generated TAF for CYYT on May 29th and 06 & 12Z worked quite well. It was an increasing southeasterly flow bringing in low stratus and fog. I believe the WIND had it going very low at 18Z while in fact it was about 19Z. This morning's (30/06Z) TAF had the visibility a bit more variable than it really was. So again we see some success in the process with stuff moving in farther in the future. However once the stuff is there, it remains to be seen what the success rate will be.

For nowcasting, persistence is hard to beat.

Verification Method

Forecasts verified using standard performance measurement method, 1

according to the average accuracy of forecasts in the 0-to 6 hour and

the 0-to-24 hour projection period of significant flying categories.

Ceiling (m) Visibility (km) Flying category

< 200 or < 3.2 below alternate

200 and 3.2 alternate

330 and 4.8 VFR

Count three

sorts of events:

1. Stanski, H., Leganchuk, A., Hanssen, A., Wintjes, D.,Abramowski, O., and Shaykewich, J., 1999: NAV CANADA's TAF amendment response time verification , Eighth Conference on Aviation, Range, and Aerospace Meteorology, 10-15 January1999, Dallas, Texas, American Meteorological Society, 63-67.

OBSERVED

YES NO

FORECAST YES hit false alarm

NO miss (non-event)

Three performance measurements calculated:

Frequency of Hits (Reliability) FOH =

False Alarm Ratio FAR =

Probability of Detection, POD =

FOH and FAR for the 0-to-6 hours are routinely tracked.

However, other more comprehensive, cost-model based schemes

would give more meaningful results in terms of forecast value. 1

Statistics

hitshits + misses

false alarmshits + misses

hitshits + false alarms

1. Forecast Verification - Issues, Methods and FAQ, http://www.bom.gov.au/bmrc/wefor/staff/eee/verif/verif_web_page.html

Statistics: Caveats

Results refer to a fully automatic and therefore handicapped system:

WIND-2 runs without guidance-improving forecaster interaction.

Results could be significantly better if WIND had forecaster input.

Statistics are summaries of statistics at these airports:

CYEG, CYFB, CYHZ, CYOW, CYQB, CYUL, CYVR,

CYWG, CYXE, CYYC,CYYT, CYYZ, and CYZF.

Each airport's statistics are given equal weight. When the statistics

for individual airports are considered, other patterns appear.

Legends in the graphs refer to 20, 30, and 40%ile, three series of

forecasts produced by WIND-2, with ceiling and visibility (C&V) based

on the 20th, 30th, and 40th percentile of C&V among retrieved analogs.

The lower the percentile, the lower the forecast of C&V.

FOH VFR, 0-6 HR, ALL SITES

80

90

100

JAN FEB MAR APR MAY JUN JUL AUG

MONTH

FO

H 20%ile

30%ile

40%ile

PER

FAR BLO ALT, 0-6 HR, ALL SITES

0

10

20

30

40

50

60

70

80

90

100


MONTH

FA

R 20%ile

30%ile

40%ile

PER

FOH ALT, 0-6 HR, ALL SITES

80

90

100


MONTH

FO

H 20%ile

30%ile

40%ile

PER

POD BLO ALT, 0-6 HR, ALL SITES

0

10

20

30

40

50

60


MONTH

PO

D 20%ile

30%ile

40%ile

PER


80

90

100


MONTH

FO

H 20%ile

30%ile

40%ile

PER


0

10

20

30

40

50

60

70

80

90

100


MONTH

FA

R 20%ile

30%ile

40%ile

PER


80

90

100


MONTH

FO

H 20%ile

30%ile

40%ile

PER


0

10

20

30

40

50

60


MONTH

PO

D 20%ile

30%ile

40%ile

PER


80

90

100


MONTH

FO

H 20%ile

30%ile

40%ile

PER


0

10

20

30

40

50

60

70

80

90

100


MONTH

FA

R 20%ile

30%ile

40%ile

PER


80

90

100


MONTH

FO

H 20%ile

30%ile

40%ile

PER


0

10

20

30

40

50

60


MONTH

PO

D 20%ile

30%ile

40%ile

PER


80

90

100


MONTH

FO

H 20%ile

30%ile

40%ile

PER


0

10

20

30

40

50

60

70

80

90

100


MONTH

FA

R 20%ile

30%ile

40%ile

PER


80

90

100


MONTH

FO

H 20%ile

30%ile

40%ile

PER


0

10

20

30

40

50

60


MONTH

PO

D 20%ile

30%ile

40%ile

PER

Conclusion

By building expert systems that combine forecaster expertise,

AI, large amounts of data (climatological and current), and

currently available computing power, we can increase forecast

quality and increase forecasting efficiency.

Acknowledgements Thesis Committee – Mohammed El-Hawary, Qigang Gao, Denis Riordan MSC Colleagues – Jim Abraham, Bill Appleby, Michel Béland, Peter Bowyer, Bill Burrows, Luc Corbeil, Daniel Chretien, Stewart Cober, Mike Crowe, Réal Daigle, Eric De Groot, Norbert Dreidger, Jack Dunnigan, Peter Houtekamer, Lorne Ketch, Alister Ling, Ted Lord, Allan MacAfee, Ken Macdonald, Martha McCulloch, Jamie McLean, Jim Murtha, Ewa Milewska, Steve Miller, Desmond O’Neill, George Parkes, Bill Richards, Steve Ricketts, Ray St. Pierre, Henry Stanski, Dave Steenbergen, Val Swail, Herb Thoms, Richard Verret, Bruce Whiffen, Laurie Wilson NRL Colleagues – David Aha, Michael Hadjimichael RAP/NCAR Colleagues – Paul Herzegh, Gerry Wiener

Future: Possible Additions and Improvements

Graphic user interface: let expert forecasters guide the data-mining to test

“what-if” weather scenarios based on various possible conditions.

Links to other software: enable WIND to help with weather watch, proactive

alerting of impending problems. For example, combine with MultiAlert to enable a

smart alert, and thus help forecasters to increase their situational awareness.

More predictors: allow data-mining to be better conditioned, e.g., duration of

precipitation, recent trends (C&V, pcpn, dp/dt), sun factors (length of day,

strength of sun), wind (back trajectory, wind run, source region, cyclonic /

anticyclonic flow), etc.

Data fusion: exploit all available data and employ data fusion techniques 1

to improve nowcasting systems, by intelligently integrating of output of various

models 2 (e.g., GEM and UMOS), forecaster input, and objective nowcasts of

precipitation (based on systems under development), and moving cloud areas

seen / detected on satellite images.

1. Intelligent Weather Systems, RAP, NCAR, http://www.rap.ucar.edu/technology/iws2. Shel Gerding and William Myers, 2003: Adaptive data fusion of meteorological forecast modules, 3rd Conference on Artificial Intelligence Applications to Environmental Science, AMS.

Future: Possible Additions and Improvements

1. Qingmin Shi and Joseph F. JaJa, 2003: Fast Algorithms for a Class of Temporal Range Queries, Proceedings of Workshopon Algorithms and Data Structures, July 30 - August 1, 2003,Ottawa, Canada. and Qingmin Shi and Joseph F. JaJa, 200?: A New Framework for Addressing Temporal Range Queries and Some Preliminary Results, submitted to Theoretical Computer Science.2. Jim Murtha, 1995: Applications of fuzzy logic in operational meteorology, Scientific Services and Professional Development Newsletter, Canadian Forces Weather Service, 42-543. Main Trend in Automation of Nowcasting: Application of Fuzzy Logic, http://bjarne.ca/ideas/trends

Faster retrieval algorithms: use reliable tree-based indexing algorithms for data

retrieval to make data retrieval 1000 times faster. 1 A faster algorithm would help

WIND to scale up and would help us to test a wider range of data retrieval

strategies, e.g., for testing what-if scenarios, forecasters could adjust conditions

with a sliding widget and see a virtually instantaneous response.

Fuzzy rule base: make WIND more of an expert system, to make it systematically

act more “intelligently”, as we learn from experts, experience, and experiments.

Add routines to deal with documented local effects and with special situations

such as radiation fog 2 and blowing snow.

Partnerships: collaborate with the Research Applications Program (RAP), NCAR

and the Aviation Weather Research Program (AWRP) to leverage limited funds,

achieve mutual benefits, and realize the above-listed improvements more quickly. 3

Hybrid Forecast Decision Support Systems

Hybrid forecast system development is a current direction of the

Aviation Weather Research Program (AWRP) 1 and the Research

Applications Program (RAP), 2 NCAR (the main organizers of

AWRP R&D).

AWRP Terminal Ceiling and Visibility Product Development Team

(PDT) project, Consensus Forecast System, a combination of:

COBEL, a physical column model 3

Statistical forecast models, local and regional

Satellite statistical forecast model

1. Aviation Weather Research Program, http://www.rap.ucar.edu/general/awrp_pmr2002

2. Research Applications Program, http://www.rap.ucar.edu

3. Cobel, 1-D model, http://www.rap.ucar.edu/staff/tardif/COBEL

Hybrid Forecast Decision Support Systems

AWRP National Ceiling and Visibility PDT research initiatives: 1

Data fusion: intelligent integration of output of various models, observational data, and forecaster input using fuzzy logic 2, 3

Data mining, C5.0 pattern recognition software for generating decision trees based on data mining, freeware by Ross Quinlan (http://www.rulequest.com), like CART Analog forecasting using Euclidean distance development of daily climatology for 1500+ continental US (CONUS) sites Incorporate AutoNowcast of weather radar in 2004-2005 4

Incorporate satellite image cloud-type classification algorithms 5

1. Gerry Wiener, personal communication, July 2003.

2. Intelligent Weather Systems, RAP, NCAR, http://www.rap.ucar.edu/technology/iws

3. Shel Gerding and William Myers, 2003: Adaptive data fusion of meteorological forecast modules, 3rd Conference on Artificial Intelligence Applications to Environmental Science, AMS.

4. AutoNowcast, http://www.rap.ucar.edu/projects/nowcast

5. Tag, Paul M., Bankert, Richard L., Brody, L. Robin. 2000: An AVHRR Multiple Cloud- Type Classification Package. Journal of Applied Meteorology: Vol. 39, No. 2, pp. 125-134.

Fuzzy LogicIntegrationAlgorithm

ProductGenerator

User

HumanInput

(> 15 min)

SelectiveClimatological

Input

Real-TimeData

Algorithms

ModelOutput

Algorithms

Data AssimilationMesoscale Model

Real-Time DataPreprocessing

QualityControl

SensorSystems

AIworkshere

Graphic UserInterface

1. RAP, Intelligent Weather Systems, www.rap.ucar.edu/technology/iws/design.htm

WeatherRadar

Nowcasts

RAP, Thunderstorm Auto-Nowcasting, www.rap.ucar.edu/projects/nowcast

Intelligent Weather Systems (RAP/NCAR) 1

“Smart Alert” Concept

Impendingproblem

Bust

0 1 2 3 4 5 6 87 9 10 11 12232221

100+603025201510987654321

Search

St. John’s

| | | | | | | | || | | | | | | | || | | | | | | | || | | | | | | | || | | | | | | | |

…| | | | | | | | |

Make

Save Send

WeatherWind

CeilingVisibilityDirectionSpeedTime…Weather

FitLoose Tight| | | | | | | | || | | | | | | | || | | | | | | | || | | | | | | | || | | | | | | | |

…| | | | | | | | |

00h 121501h 131402h 1412...12h 1408

00h R-L-01h R-L-02h L-...12h L-

Search

AMD TAF CYYT 270010Z 270024 1315KT 2SM -RA BR OVC006 TEMPO 0002 1/2SM -DZ FG OVC003 FM0200Z 14010KT 1/2SM -DZ FG OVC002 TEMPO 0224 1/4SM -DZ FG OVC001 RMK NXT FCST BY 06Z=

Search Make

INTEGRATION

CLIMATEARCHIVE

data

PRODUCTDISPLAY(editable)

HEADS-UPALERT &DISPLAY

ACTUALWEATHER

MAP(animated)

GUIDANCEDISPLAY(satellite,

NWP, etc.)

FORECASTER(interacts, intervenes)

awareness and knowledge

PREDICTION

UPPER AIR

SATELLITE

METAR

REAL-TIMEOBSdata

RAW, QC’dWEATHER

data

MODELLEDWEATHER

NWPdata

PRODUCTGENERATION

PRODUCTSinformation

MODELLEDWEATHER

MAP(editable)

DSS(interaction withintegration and

prediction)

PRODUCTSPECIFICATIONS

CONSISTENCYCHECKING

TRANSLATION

FORECAST

EXTRAPOLATION

PROJECTEDOBS

AIknowledge

USER

MODEL-BASEDWEATHERELEMENTS

VERIFICATION

0 time

official forecast

actual trend

!

Graphic interventionFirst resort

Direct interventionLast resort

data and information• up-to-the-minute intelligent data fusion• abstract features• derived fields• intelligently composed “interest fields”

RADAR

DAdata

information• special interests• cost-based decision-making models

DECISION SUPPORT SYSTEMS *

POST-PROCESSING

Battleboard raises forecaster’s situational awareness

GUI leverages forecaster’s actions

* Forecaster Workstation User Requirements Working Group meeting notes, 2000: Decision support systems for weather forecasting based on modular design, updated slightly for Aviation Tools Workshop in 2003.

Decision Support Systems Design

Generic: no-name, conceptual design that could link and

integrate the most useful elements of WIND, AVISA, MultiAlert,

SCRIBE, FPA, URP, and so on in evolving WSP application, NinJo.

Modular: shows where distinct sub-tools / agents can be developed.

Working in this way, individual developers could work on isolated

sub-problems and anticipate how to plug their results into a larger

shared system. As technology inevitably improves, improved modules

can be easily installed and quickly implemented.

User-centered: forecast decision support systems from forecaster's

point of view, designed to increase situational awareness.

Hybrid: combines complementary sources of knowledge, forecasters

and AI, to increase the quality of input data and output information.

Intelligent integration of data, information, and model output, and

use of adaptive forecasting strategies are intrinsic in this design.

A graphical illustration to fuzzy logic, http://www.mcu.motsps.com/lit/tutor/fuzzy/fuzzy.html

Since we can assign numeric values to linguistic expressions, it follows that we can also combine such expressions into rules and evaluate them mathematically.A typical fuzzy logic rule might be:

If temperature is warm and pressure is low then set heat to high

Fuzzy logic


How Rules Relate to a Control Surface

A fuzzy associative matrix (FAM) can be helpful to be sure you are not missing any important rules in your system. Figure shows a FAM for a control system with two inputs, each having three labels. Inside each box you write a label of the system output. In this system there are nine possible rules corresponding to the nine boxes in the FAM. The highlighted box corresponds to the rule:

If temperature is warm and pressure is low then set heat to high


The input to output relationship is precise and constant. Many engineers were initially unwilling to embrace fuzzy logic because of a misconception that the results were not repeatable and approximate. The term fuzzy actually refers to the gradual transitions at set boundaries from false to true.

Three Dimensional Control Surface

CBRInference Engine

CBRKnowledge Base

Problem Input

Assign Indices Indexing Rules

Case Base Input + Indices

Case Retrieve MatchMemory Rules

Retrieved CaseStore

Adapt AdaptationAssign Indices Rules

Proposed SolutionNew Case Test New Solution

Solution Failure Description Repair Repair

RulesExplain Causal Analysis

PredictiveFeatures

1. Adapted from Riesbeck and Schank 1989

difficult problem

potential endless loop

CBR needs methods for acquiring domain knowledge for retrieval and adaptation.

Classic CBRFlowchart 1

Documents

Analog forecasting of ceiling and visibility using fuzzy logic and data mining Bjarne Hansen Meteorological Research Branch Meteorological Service of Canada