From: AAAI-94 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved.

Case-Based Reasoning for Weather Prediction

C. Vasudevan
AUV Group, Department of Ocean Engineering,
Florida Atlantic University,
500, NW 20th St., Boca Raton, FL 33431.
email: vasu@transquest.oe.fau.edu

Abstract tent, wind direction and speed, sea and land surface
temperatures, location and density of heat islands, at-
Computer-based forecasting of weather was first exper-
mospheric pressulre, and water vapor content. In ad-
imented in 1950 at Princeton University. Since then, dition to matching individual records, the variational
there have been newer and more accurate methods to patterns of these factors are also to be considered in
predict the incoming climate. One common practice assessing the situation. There are a number of rules
of weather prediction is by using the general circula- of thumb derived from past experience that should
tion models which are based on the laws of physics be reflected in the case indices. Some examples are:
(J.M.Moran & M.D.Morgan 1986). These models are Falling air pressure may indicate the approach of a
highly complex and computational intensive limiting stormy weather. A wind shift from northwest to west
their use for only short range predictions and that to southwest is usually accompanied by warm air ad-
too needing supercomputers. The accuracy of fore- vection. Building the casebase for weather prediction
casting deteriorates rapidly for periods longer than 48 benefits from the archives already maintained by vari-
hours and it often becomes minimal beyond 10 days ous meteorological agencies. The ‘repair’ or adaptation
due to imperfections in the models. The analog tech-
of matching case(s) is based on the immediate past in-
nique of weather forecasting is another approach which formation and the interpretation of mismatches in the
searches for periods in the past when the current condi-
observed data with the past case data.
tions were similar and use the past spatial patterns as
The case-based reasoning scheme naturally suits
analogs (J.T.Houghton, G.J.Jenkins, & J.J.Ephraums weather prediction based on historical records. Al-
1990). Long term trends and recurring events guide
though recurrence of climatic patterns may not be al-
the decisions. This is more relevant for long range pre- ways dependable, this approach provides a simpler and
dictions as well as in single station predictions. The faster method to predict the future weather conditions
araudog method is relatively simple compared to the than the complex numerical models. Moreover, this
complex processes of development, validation, use, and scheme establishes a baseline for a detailed and more
maintenance of numerical models. accurate scheme based on the physics of the system.
The analog technique closely resembles the princi- This project is in a preliminary stage and a detailed
ples of case-based reasoning (CBR) (Kolodner 1993; knowledge engineering of weather patterns and predic-
Hammond 1989). The CBR scheme attempts to iden- tion skills is needed to build the case library.
tify a solution by searching a historical database of
solutions. Rather than performing statistical compu- References
tations on past records, CBR attempts to retrieve one
Hammond, K. 1989. Case-Based Planning: Viewing
or a few best matching cases from its casebase and
Planning as a Memo y Task. Academic Press.
modifies them to fit the current scenario. The CBR
approach often results in a faster synthesis of solutions J.M.Moran, and M.D.Morgan. 1986. Meteorology.
compared to rule-based reasoning or reasoning from Macmillan Publishing Company.
first principles. J.T.Houghton; G.J.Jenkins; and J.J.Ephraums. 1990.
This paper discusses the relevance of CBR in Climate Change: The IPCC Assessment. Cambridge
weather forecasting and outlines an indexing and eval- University Press.
uation scheme for prediction. The major issue is the Kolodner, J. 1993. Case- based Reasoning. Morgan-
task of identifying a set of indices to retrieve match- Kaufmann.
ing past case records and interpret them in the current
context. For weather cases, the indices are typically
the observations of relevant atmospheric parameters
such as cloud amount, cloud altitude, cloud water con-

1494 Student Abstracts