Anda di halaman 1dari 5

Using ANFIS and Fuzzy Clustering to Predict

Extreme Climate in Indonesian regions


The Houw Liong2) and Plato M.Siregar1)

ICICI 2007 proceedings


Jl. Ganesha 10
Bandung, 40132 INA
1
) Faculty of Earth Science and Mineral Technology, ITB
2
) Faculty of Mathematics and Natural Sciences, ITB

Abstract The relative positions of the sun in the sky during the
seasons, as well as the cycles of solar activity influence the
Extreme climate in Indonesia is influenced by four weather and climate throughout the Indonesian archipelago.
main quasi periodic cycles: Solar Activity Cycle (Sunspot Solar irradiance and ultraviolet intensity increases with higher
Numbers Cycle), Galactic Cosmic Ray Cycle, El Nino solar activity. This in turn will be followed by coronal mass
Southern Oscillation (ENSO) Cycle, and Indian Ocean Dipole ejection (CME) that increases the charged particles emitted by
Mode (IOD) Cycle. It can be shown that solar activity cycle the sun which could alter the interplanetary magnetic field, and
can be considered as primary cycle that influence other cycles. hence the intensity of galactic cosmic rays reaching the earth.
In practice eastern Indonesian region is dominantly The galactic cosmic ray intensity reaching the earth decreases
influenced by ENSO. When the heat pools moves to eastern with higher solar activity. Thus the solar activity is often
Indonesian region, then rainfall in this region will be above considered as the dominant factor that determines the dynamics
normal. On the other hand when the heat pool leaves eastern of climate [1, 2]. The dynamics of earth's atmosphere and
Indonesian region and moves to Pacific Ocean then the rainfall oceans, evaporation, clouds formation and rainfall, are
in this region will be below normal. influenced by the solar energy entering the earth. Several
During a typical Indian Ocean Dipole Mode (IOD) studies indicate that strong correlations exist between the cloud
event the weakening and reversal of winds in the central cover and the intensity of galactic cosmic ray reaching the
equatorial Indian Ocean lead to the development of unusually earth [3].
warm sea surface temperatures in the western Indian Ocean. During 1645 – 1715 exceptionally low solar activity
IOD negative means wet condition or the rainfall will be above (also known as the Maunder minimum) which means high
normal along the western Indonesian region. intensity of galactic cosmic ray reached the earth increased
Precipitation in Pontianak region which represent cloud cover that led to low temperatures causing what is known
middle Indonesian region correlated strongly with sunspot as the little ice age.
numbers cycle (solar activity cycle). The present study shows that there is a strong correlation
Using ANFIS (Adaptive Neuro Fuzzy Inference System) between rainfall in the Indonesian archipelago and solar
we are able to predict sunspot numbers cycles so that extreme activity or sunspot numbers.
weather in Indonesian regions can be predicted.
Fuzzy c-means is used to classify regions that are II. THE RELATION BETWEEN SOLAR ACTIVITY
influenced strongly by sunspot numbers (solar activity), IOD, AND CLIMATE
and ENSO cycles. This method is based on fuzzy set as fuzzy The galactic cosmic rays collide with air molecules in
c-partition of three cycles above and as cluster center. Fuzzy c- the upper atmosphere and produce secondary particles.
partition matrix for grouping a collection of n data set into c Generally the charged particles so produced cannot penetrate to
classes. lower layers of the atmosphere, except gamma ray, neutrons
This study explores the physical of climate predictions and the muons. When gamma ray, neutrons and muons interact
and classifications of Indonesian regions and its physical with the air molecules or water molecules, they become
interpretations. charged and together with aerosols particles act as
Keywords: ANFIS, fuzzy clustering, climate, solar activity condensation nuclei for the formation of clouds. The cosmic
ray becomes the source of ions in the air besides radiation
I. INTRODUCTION coming from earth originated by the radio isotope radon.
During the sunspot minimum, the intensity of the membership function in this node and can be any appropriate 
galactic cosmic ray that penetrates earth atmosphere becomes functions that are continuous and piecewise differentiable such 
maximum which in turn increases the coverage of clouds. This
as Gaussian , generalized bell shaped, trapezoidal shaped and 
implies that solar irradiation reaching the earth will be
minimized. Conversely, during solar activity maximum or triangular shaped functions. Assuming a bell shaped function 
sunspot maximum, the intensity of galactic cosmic ray as the membership function, Ai can be computed as, 
reaching lower levels of the atmosphere decreases, less cloud
condensation nuclei are produced, hence the cloud cover 1
µA ( x ) =
decreases, furthermore extra energy received from flares during x −c i
2b
(2)
prominent eruptions, maximizes the amount of solar energy 1+
reaches the earth. ai
Although global cloud cover produces a warming effect
or the greenhouse effect, but a cooling effect due to reflections
against direct solar irradiation is more dominant factor [1, 3]. ai  and ci  are   the   parameter   set.   Parameters   in   this   layer   are 
Furthermore during solar activity maximum, the
referred to as premise (antecedent) parameters. 
intensity of ultraviolet that penetrates the earth increases. Solar
activity maximum usually is followed by increasing coronal Layer 2: Every node in this layer is a fixed node labeled  Π  , 
mass ejection. Both effects caused greater amount of energy which multiplies the incoming signals and outputs the product. 
penetrates the earth and this will influence the climate through For instance, 
the dynamics of the atmosphere and oceans.

O 2,i = w 1 = µ Ai (x ) µ Bi ( y) i = 1,2 (3)

III. Adaptive Neuro­Fuzzy Inference System (ANFIS)
Each node output represents the firing strength of a rule. 
and Fuzzy Clustering
Layer 3: Every node in this layer is a circle node labeled N. The 
ith node calculates the ratio of the  ith rule's firing strength to 
Adaptive network­based fuzzy inference system used a 
the sum of all rule's firing strengths. Output of this layer will be 
feed forward network to search for fuzzy decision rules that 
called normalized firing strengths. 
perform well on a given task. Using a given input­output data 
set   ANFIS     creates   a   fuzzy     inference   system   whose 
wi
membership   function   parameters   are   adjusted   using   a  O 3, i = w i = , i = 1,2
w + w (4)
backpropagation   algorithm   alone   or   combination   between   a  1 2

backpropagation algorithm with a least squares method. This 
Layer 4: Node i in this layer compute the contribution of the ith 
allows the fuzzy systems to learn from the data being modeled. 
rule   towards   the   model   output,   with   the   following   node 
ANFIS provide a method for the fuzzy modeling procedure to 
functions: 
learn information from the data set, followed by creating the 
membership function parameters that best performing the given 
O 4,i = w i f i = w i (p i x + q i y + ri ) (5)
task. Consider a first order Takagi­Sugeno fuzzy model with a 
two input, one output system having two membership functions 
Layer 5: The single node in this layer is a fixed node labeled 
for each input.[6] 
that   computes   the   overall   output   as   the   summation   of   all 
The functioning of ANFIS is a five layered feed forward neural 
incoming signals.
structure and the functionality of the nodes in these layers can 
Overall output = 
be summarized as follows: 
Layer 1: Every node i in this layer is an adaptive node with a  ∑i w i f i
node output defined by: O 5 = ∑ w if i =
i ∑ w i (6)
i

O = µ (x i = 1,2 ,
1 ,i A i
), for or
O = µ (y i = 3 , 4 ,    (1) Using ANFIS and data from Royal Observatory of Belgium
1 ,i B i − 2
), for
and Sunspot Index Data Center http://www.astro.oma.be/SIDC
where x(or y) is the input to the node; Ai (or Bi­2) is a fuzzy set  we can get prediction of sunspot numbers time series as in
associated with this  node,  characterized  by the  shape  of the  Figure 1.
We use rainfall data in Indonesia from NCEP
Reanalysis at
http://www.cdc.noaa.gov/cdc/data.ncep.reanalysis
and fuzzy c-means method with three seeding regions
for initial matrix fuzzy c-partition of three cluster centers i.e. :

ANFIS PREDICTION

200
Numbers Sunspot

150

100 Figure 2. Climate regions in Indonesia based on fuzzy clustering


50

0
IV.The Correlation of Sunspot Number to Rainfall in
Indonesia
1948

1952

1956

1960

1964

1968

1972

1976

1980

1984

1988

1992

1996

2000

2004

2008

2012
Years
With the equator crossing Indonesia, the sensible heat
ANFIS Prediction Obs. Sunsspot
flux plays an important role in global circulations. The latent
heat which originates mainly from the release of latent heat
Figure 1. ANFIS prediction of sunspot number when water vapor condenses into clouds droplets(a number of
time series. large clouds form through convections in the Inter Tropical
Convergence Zone (ITCZ) which is above Indonesia). The cold
Bukittingi region for western Indonesia,, Jayapura monsoon season in northern hemisphere (Asian monsoon) and
region for eastern Indonesia and Pontianak region for middle in the southern hemisphere (Australian monsoon) are
Indonesia. This method is based on fuzzy set as partition influenced by the heat source distribution or the release of
matrix for grouping a collection of n data set in to c classes, we latent heat above Asia and in the neighborhood regions. At
define object function for fuzzy as Euclidian distance. The present it seems that the Indonesian zone holds the key to
result of the clustering is shown in Figure 2. It is shown that southern oscillation system which determines the forcing of
Jakarta region (Jabodetabek) is similar (0.6) to middle region
which is dominated by solar cycle, is similar (0.5) to western
region which is dominated by IOD cycle, and is similar (0.5) to Jaya Pura
eastern region which is dominated by ENSO cycle. The result
of climate clustering in Indonesia is using the following 400 200
algorithm. 300 150
200 100
Fuzzy c-means Algorithm 100 50
0 0
1. Fix c (2≤c≤ n) and select a value for parameter m’
,initialize the partition matrix U(0), each step in this Y ears

algorithm will labeled r, where r=0,1,2,..


Avg precip sspot
2. Calculate the c centers {vi(r)} for each step.
3. Update the partition matrix for rth step, U (r) as follow:
−1

 (r ) 
2 /(m'−1) 
    Figure 3 Yearly precipitations in Jayapura region vs.
 c d 
µ (r +1) = ∑  ik 
  for I = Φ (7) sunspot numbers.
ik  j =1 (r )   k
  d jk  
   

4. If U
( r +1)
− U ( r ) ≤ ε ,stop; otherwise set r=r+1 and
return to step 2
5. Calculate the similarities of the c center :
n m'
∑ µik .x
kj
vij = k =1n
m'
∑ µik
k =1 (8)
Pontianak Region Jabodetabek
Correlation Sunspot vs Precip =0.88
400 200

mm/month
Sunspot/Pre

200 200
200 100
cip

100 0
0 0
0 -200
Years
Years
Avg Precip Avg-sspot
ave-sunspot ave-precip

Figure 5. Maximum precipitation in Jabodetabek in


Figure 4. Correlation between sunspot numbers and 1968, 1981, 1992, 2002 correspond to sunspot number
yearly precipitation in Pontianak region. maximum. Precipitation maximum in 1976, 1986, 1996,
El Nino Southern Oscillation (ENSO). Therefore, and (2007) correspond to sunspot number minimum or
Indonesia, through which the equator crosses, has the galactic cosmic ray maximum.
maximum sensible heat flux, high rainfall, and monsoon After 1976 sunspot numbers maximum SMax and
circulations. Consequently, it is one of the most primal zones sunspot numbers minimum SMin correspond to precipitations
for convection processes, an equatorial-tropical zone where above normal also to La Nina and maximum eruptions ME or
Coriolis effects are practically nullified, where atmospheric CME corresponding to precipitations below normal and also to
circulations are very different compared to the extra-tropical El Nino.[4, 5] In Pontianak region which represent middle
zones. Indonesian Maritime Continent, the yearly precipitation is
The observations and studies on Indonesian climate are mainly determined by sunspot cycles (Figure 4). Precipitations
limited, and the mathematical formulations of tropical above normal occur at sunspot maximum SMax, and
dynamics are far more complex relative to those in the extra- precipitations below normal at sunspot minimum SMin.
tropical zones. The distinct daily convection variability induced Precipitations in east Indonesia which represent North
by land-sea wind circulations over some islands in Indonesia Australia Indonesian Monsoon are influenced by ENSO similar
characterizes the aspect of rainfall throughout the Indonesian to those observed in Jayapura region. (Figure 3) Precipitations
Archipelago which are very different from other regions on the in Jakarta region or Jabodetabek are weakly influenced by
earth. The studies mentioned above, show that rainfall is an ENSO. The peaks of yearly precipitations correspond to the
important quantity in the Indonesian Archipelago and sunspot peaks of sunspot numbers, but at the sunspot numbers
is believed to be the major predictor. minimum which correspond to galactic comic ray maximum,
Using rainfall data in Indonesia from NCEP Reanalysis the yearly precipitations also maximum.(Figure 5).
http://www.cdc.noaa.gov/cdc/data.ncep.reanalysis The west Indonesian region is mainly influenced by IOD
We can get links between sunspot numbers and rainfall, that also correlated to solar cycle. [5]
although the correlations which existed in general are weak. In The fuzzy c-means clustering shows that the western
other words, these signify that the dynamics of the atmosphere Indonesian region is influenced mainly by IOD, the eastern
is being viewed as the cause of the small correlations. Indonesian region is influenced mainly by ENSO and the
However, in the case of static model atmosphere, determination middle region is mainly influenced by solar activity.
of correlations based on data averaging of sunspot numbers on So, by knowing sunspot number time series as predicted
yearly basis against the yearly rainfall for various regions in by ANFIS and fuzzy clustering of climate regions we can
Indonesia, one comes to time series as shown in Figures 3, 4 predict the coming extreme weather for each regions in
and 5 . Indonesia
From Figure 3 we can conclude that eastern Indonesia REFERENCES
(Jayapura region) which represented Eastern Indonesian [1] H. Svensmark, Cosmoclimatology : a new theory emerges. Astronomy &
Geophysics,Vol. 48, pp 1.18-1.24, 2007
Maritime Continent is strongly influenced by ENSO. [2] T. Landscheidt, Solar Activity: A Dominant Factor in Climate Dynamics,
Schroeter Institute for Research in Cycles of Solar Activity,
http://www.johndaly.com/solar/solar.htm, 1988.
[3] K.S. Carlslaw, R.G. Harrison, J. Kirkby, Cosmic Rays, Clouds, and
Climate, Science’s Compass, Vol. 298, 2002.
[4] T. Landscheidt, New ENSO Forecast Based on Solar Model, Schroeter
Institute for Research in Cycles of Solar Activity, 2003.
[5] The H. L., P. M. Siregar,Using System Dynamics of Ciliwung River to
Predict Floods, Workshop on Nonlinearity 2k6, IPB, Bogor, 2006.
[6] J.S.R.Jang, C.T. Sun, E. Mizutani, Neuro-Fuzzy and Soft Computing,
Prentice Hall, Inc., 1997.

Anda mungkin juga menyukai