IOP Conference Series: Earth and Environmental Science

PAPER • OPEN ACCESS

Forecast of Tropical Cyclone Occurrences based on Fuzzy Logic

Algorithm
To cite this article: A M Irawan et al 2019 IOP Conf. Ser.: Earth Environ. Sci. 303 012044

View the article online for updates and enhancements.

This content was downloaded from IP address 125.166.184.113 on 10/10/2020 at 10:03

International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

Forecast of Tropical Cyclone Occurrences based on Fuzzy Logic

Algorithm

A M Irawan 1*, H A Nugroho 2, P P Simanjuntak 1 and S I Sugiarto 1

1
Department of Climatology, School of Meteorology Climatology and Geophysics
(STMKG), Jakarta, 15221, Indonesia
2
Instrumentation Laboratory, Department of Instrumentation, School of Meteorology
Climatology and Geophysics (STMKG), Jakarta, 15221, Indonesia
amir.irawan@stmkg.ac.id
Abstract. Most studies of the tropical cyclone (TC) forecasting are focused on the track and the
wind radius forecasting, even though a tropical cyclone formation alerts also developed. This
paper takes a different approach and explores and forecast the TC occurrences. This study
presents the development of fuzzy logic (FL) models for predicting the occurrence of TC from
five primary TC genesis parameters. These parameters are low-level relative vorticity (θ),
horizontal wind of upper troposphere (u), sea surface temperature (SST), equivalent potential
temperature (𝜃𝑒), and specific humidity (q). The FL model was developed by employing the
trapezoidal and triangular fuzzy membership functions for the input and output variables. The
fuzzy rules were inferred from the TC genesis parameters data, with a daily calculation period
from 1989 to 2018. The amount of TC genesis parameters when the cyclone occurred of the
lowest, middle, and upper tercile reconstruction was used as a threshold to build FL model. The
model satisfactorily simulated the occurrence of TC with comparable error measures. The result
exhibits the accuracy at 0.75 (range: 0 to 1, perfect score: 1). The evidence shows that the result
provides insights into the adequacy of FL methods for forecasting the TC occurrences.

Keywords : tropical cyclone, fuzzy logic, relative vorticity.

1. Introduction
TC is the generic term for a non-frontal synoptic scale low-pressure system originating over tropical or
subtropical waters with organized convection and definite cyclonic surface wind circulation. Tropical
cyclones with maximum sustained surface winds of less than 17 ms-1 are generally called "tropical
depressions (TD)." Once a tropical cyclone achieves surface wind strengths of at least 17 m/s, it is
typically called a "tropical storm" or "tropical cyclone" and assigned a name [1]. In the northern
hemisphere, TC occurs between June and November peaking in September; moreover, in the southern
hemisphere, the season lasts from November to April. More than one TC can occur in the same ocean
and region at once. Among cyclones that occur in the Southern Hemisphere, almost half are formed
above the Northern coast of Australia, a quarter appear above the South Pacific Ocean, and others occur
above the ocean of South Indonesia [2]. Tropical cyclones are perhaps the most devastating of natural
disasters both because of the loss of human life they cause and the large economic losses they induce
[3-7].
One of the most challenging subjects in term of ocean-atmosphere interaction is that of understanding
the tropical cyclone (TC) occurrences. Many opinions stated that the environmental conditions and the
physical mechanisms which bring about tropical cyclone formation is important to understand why and
how TC form. Among scientists who have studied this phenomenon there lies a wide variety of opinion.
For instance, Neiburger, Edinger and Bonner [2] observed that TC possibly form when sea surface

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

temperatures (SST) depicts the magnitude more than 27°C and do not form between the 4°N and 4°S
(tropical areas) of the equator. Based on this TC form requirements, TC will be less likely to pass through
Indonesia due to a tropical characteristic of Indonesia region. However, the effects of TC that occur
around Indonesia could affect bad weather and meteorological disaster in various places in Indonesia.
Vulnerability to tropical cyclones is becoming more pronounced because the fastest population growth
is in tropical coastal regions. In Indonesia, for example, TC Cempaka that occurred on 28 October 2017
generated extreme weather has caused floods, landslides, and tornadoes in 21 regencies/cities in Java
and Bali. Temporary data collected by National Agency for Disaster and Management (BNPB) post, the
disaster occurred in the large area of Java such as Semarang, Kulon, Progo, Ponogoro, Bantul, and
Kudus [8]
In this research, fuzzy logic (FL) algorithm was used to forecast both TC and TD occurrences. This
method has some advantage, i.e. (i) it mimics the human thinking and reasoning, (ii) it can detect the
uncertainties of dynamic system behaviors, (iii) it is based on training and experiences rather than of the
theory. This new approach is necessary since it could have some implications for the TC and TD
forecast.

2. Data and Method

2.1 Study Site
To assess TC forecasting related to the main parameters in the formation of TCs such as wind—pressure
relationship, finding a suitable study site is important [9]. Tropical cyclones generally occur in the
tropics with low latitudes between 10° and 20° N / S from the equator [10]. In addition, the most active
TC growth area for the Southern Hemisphere is Australia's northern coast which experienced 124
incidents within 100 years [11]. As explained by the Australian Bureau of Meteorology (BOM), this
region reaches an average of 10 times of TC occurrences per year. From this literature, Australia's
northern coast is an area that has higher TC occurrences more than many places in the world. In the
Southern Hemisphere, almost TC occurrences are formed above the northern Australia ocean, South
Indonesia and South Pacific Ocean [2]. Therefore, in this study, the Southern Ocean of East Nusa
Tenggara (15°-10°S and 115°-130° E), Indonesia, is employed as the study site. This area is an example
of a populated TC occurrence that is highly vulnerable to the storm which is mainly caused by TC.

Figure 1 Domain of the study

As illustrated in figure 1, the layer of the study domain shows an ocean between East Nusa Tenggara
and Australia North coast. Based on the archive of tropical cyclones [12], there were 56 occurrences of
tropical cyclones in the research grid during the period 1989-2018 in winter and spring periods.

2
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

2.2 Input Selection

Gray [13] explained that the frequency of tropical cyclones could be directly related to a combination
of several physical parameters called the main parameters in the formation of tropical cyclones. These
parameters are:
1. The amount of vorticity in the lower troposphere. Tropical cyclones only form in large areas of
vorticity in the lower troposphere.
2. The zonal wind vertical shear role. According to Gray [13], the requirements of tropical cyclones
are a weak zonal wind shear approximately at 0-1 m/s.
3. The role of ocean thermal energy. Tropical cyclones have a large influence on sea surface
temperature, when the sea surface temperature changes, it changes the cyclone characteristics.
Tropical cyclones can have a profound influence on the temperature of the ocean over which they
travel. The altered ocean temperature, in turn, can feedback and alter the character of the cyclone.
4. The influence of the surface to the equivalent potential temperature in the middle troposphere.
Cyclone formation is related to moist buoyancy potential or the extent of the layer boundary of the
atmosphere to the middle troposphere. Representative Buoyancy to study TC characteristics is the
difference between the surface (1000 mb) and 600 mb layer.
5. The role of moisture in the middle troposphere. Moisture parameters vary from 0 to 1. The
development of cyclones does not occur if the humidity in the layer 500-700 mb is less than 40%.
This factor increases linearly by 1 in the humidity between layers 700-500 mb reaching 100%, or
humidity parameter is between 40% and 70%. So tropical cyclones only form in regions with
relatively high humidity (RH) or specific humidity (q) in the middle troposphere.

Therefore, the input parameters used in this study are the main parameters in the formation of tropical
cyclones, i.e., low-level relative vorticity (ϑ), the horizontal wind of upper troposphere (u), sea surface
temperature (SST), equivalent potential temperature (θe), and specific humidity (q). The input variables
are selected during the rainy season in Indonesia on December, January, February (DJF) and March,
April, May (MAM). These periods were selected due to the highest frequency of TC occurrences in the
south of Indonesia started from October to May with a peak in December to March [14].
All of the input parameters are obtained from the Japan Meteorological Agency (JMA), Tokyo,
Japan [15]. Furthermore, this study employed the input parameter data with a daily calculation period
from 1989 to 2010 as a training data and 2011 to 2018 as a validation data.

2.3 Fuzzy Logic (FL) Reconstruction

Bellman and Zadeh [16] suggested that fuzzy is a different aspect of uncertainty with randomness. Then
they proposed a mathematical form to see how obscurity can be expressed in a human language whose
approach is called "fuzzy logic (FL)". In fuzzy it is known that membership degrees have intervals [0,1].
With fuzzy logic, we can express the same information (less specific), then manipulate it, and conclude
from that information. FL is used in this study because it can handle the determination of event
predictions based on input data.
Many scientists have employed the FL algorithm to meteorology and hydrology problems [17-20].
The method of FL algorithm is to consider the system depicts in the form of subsets or fuzzy approaches,
each of which is stated with words such as “low”, “medium”, “high”, etc. An FL algorithm consists of
parameters that have varying membership degrees in the set. This idea is in contrast with classical
algorithm theory because for crisp sets an element either is a member of that set or is not. FL rule-based
on systems can be used as useful representations of either simple or complex physical problems. A small
number of fuzzy sets leads to unrepresentative forecasts whereas a large number leads to many
calculations. In this research, we set 3 number of fuzzy and labeled with, low, medium, and high. This
study depicts the following steps to construct an FL algorithm. These are:
(i) Fuzzification of the input and output parameters by considering convenient linguistic subsets, i.e.,
low, medium, high
(ii) Fuzzy rules construction based on IF-THEN rules to simulate the model. The rules relate the
combined linguistic subsets of input parameters to the output fuzzy sets. The input subsets within

3
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

the premise part are combined interchangeably with the logical “and” or “or” conjunction whereas
the rules of this study used a logical “and” conjunction based on the formation of tropical cyclones,
(iii) The implication part of a fuzzy system is defined as the shaping of the consequent based on the
premise (antecedent) part, and finally
(iv) The result is a fuzzy set, and therefore, requires defuzzification to arrive at a crisp value, which is
required to simulate the model.

In the applications of the FL algorithm in control and forecasting, there are mainly two approaches:
the first one is the Mamdani method, and the other is Takagi-Sugeno. For the Mamdani method, they
use clear procedures, i.e., fuzzification, logic decision, and defuzzification procedure. For Takagi-
Sugeno algorithm [21], however, does not have a clear defuzzification method. For the Mamdani
approach, the outcome of each IF-THEN rule will be a fuzzy set for the output variable so that the step
of defuzzification is indispensable to obtain the crisp value of the output variable. Therefore, this study
utilising Mamdani procedure to construct FL rules.

2.4 Tropical Cyclone (TC) Forecasting Model

Before determining the threshold of the fuzzy membership function, this paper aims at examining the
lagged input data to predict TC and TD occurrences. We used the cross-correlation method to evaluate
the relationship between the lag time of each input variable and both TC and TD occurrences. The lag
time with best biserial correlation with TC and TD occurrences were chosen as predictors.
Furthermore, we categorized the predictor threshold based on 33th percentile of TC and TD
occurrences for 1991-2010 periods and divided into three classes (low, moderate & high). After the
fuzzy rule base is identified and a defuzzification algorithm is selected, forecast accuracy is tested using
a different set of the historical data set (training data) from the one used to obtain the rule base. If it is
unsatisfactory, then the number of fuzzy membership functions and/or shape of the fuzzy membership
functions can be changed, and a new FL rule base is obtained. The iterative process of designing the
rule base, choosing a defuzzification algorithm, and testing the system performance may be repeated
several times with a different number of fuzzy membership functions and/or different shapes of fuzzy
memberships. The FL rule base that provides the minimum error measure for the test set is selected for
real-time forecasting.
This study used trapezoidal and triangular membership function in term of the main characteristics
of five TC and TD input parameters. The results of the model will be verified using 'categorical statistics'
from the 'contingency table'. Contingency table or also called cross-tabulation table or crosstab is a table
arranged based on data tabulation according to 2 or more categories that are displayed because an
element with other elements is in conformity or relationship. Each column of the matrix is an example
in the prediction class, while each row represents an example in the actual class. Dimensions or sizes of
various contingency tables, the simplest are dimension 2, like the example below:

Table 1. Contingency Table Form

Observed
Forecast Total
Yes No
Yes Hits False alarms Forecast yes
No Misses Correct negatives Forecast no
Total Observed yes Observed no Total

The four combinations of forecasts (yes or no) and observations (yes or no), called the joint
distribution, are hits (event forecast to occur, and did occur), misses (event forecast not to occur, but did
occur), false alarm (event forecast to occur, but did not occur) and correct negative (event forecast not
to occur, and did not occur).

4
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

𝐻𝑖𝑡𝑠 + 𝐶𝑜𝑟𝑟𝑒𝑐𝑡 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 (𝐴𝐶) = (1)
𝑇𝑜𝑡𝑎𝑙

According to Provost and Fawcett [22], performance measurement algorithms that can be measured
include Accuracy (AC) which is the correctness of the whole model and is calculated as the sum of the
correct classifications divided by the total number of classifications. This skill was calculated based on
the proportion of the correct number of predictions and is determined using the equation of the
contingency table above.

3. Results and Discussion

3.1 Time Lag Analysis and Fuzzy Rules Construction
Figure 2 shows the relationship between the lag time of each model input with the occurrence of TC and
TD. All of the inputs exhibit the highest correlation in lag 1, and those relationships decrease with an
increase of lag time.

Figure 2. Biserial correlation between input parameters and TC and TD occurrences

Among input parameters, negative relationships were found in low-level relative vorticity (ϑ), the
horizontal wind of upper troposphere (u), sea surface temperature (SST), and equivalent potential
temperature (θe).

In contrast, specific humidity (q) employed positive correlation with TC and TD occurrences in the
south of East Nusa Tenggara. Based on the above time lag analysis, we used input variables based on
the highest correlation with TC and TD occurrences. Furthermore, we construct FL rules regarding the
input parameters that were chosen. The class of FL rules was a construct based on 33th percentile of TC
occurrences data. Table 2 exhibits the threshold value of each input parameter in each class.

Table 2. The threshold values of input parameters were divided into three classes based on the lowest,
medium, and highest tercile of TC occurrences from 1991 to 2010.

Θ U SST θe Q
(10 , S-1)
-6
(ms-1) (oC) o
( C) (10 , gkg-1)
-4

Low -31 to -8 -18.75 to -8.64 26.78 to 28.18 7.39 to 12.20 8.57 to 41.56
Moderate -8.1 to 0 -8.65 to -3.55 28.19 to 29.09 12.21 to 14.10 41.57 to 49.55
High >0 -3.56 to 14.56 >29.09 >14.10 >49.55

After the FL rule base is identified and a defuzzification algorithm is selected, forecast accuracy is
tested using a different set of the historical data set (training data) from the one used to obtain the rule
base. FL rules were obtained and changed based on the best accuracy of training data. From Table 2, we
build 243 rules for the occurrence of TC and TD. We set the rules regarding the highest accuracy of

5
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

training data from 1991 to 2010. It should be noted that the FL approach will work reasonably as the
degree of fuzziness is suitable with the variables and the natural behavior of the system. Furthermore,
FL rules are constructed between input parameters in IF-THEN format regarding the variables and
expert decision. Lastly, results are defuzzified to a specific number as an output thought to the best
represent fuzzy (figure 3).

Figure 3. The instance of FL model reconstruction (rule 1) by using centroid method for ϑ= -9.5 S-1, u= -
16 m/s, SST= 29.5oC, θe=15oC, and q= 39.5 g/kg. In this study, we build and simulate 243 FL rules that
depict the complexity of the model.

Many scientists used centroid as defuzzification method to model FL algorithm [23]. Therefore, in
this study, we continue to use the centroid method to compute FL algorithm of TC and TD occurrences
based on the expert decision. Figure 3 was made to facilitate understanding of the relationship between
the input parameter and both TC and TD occurrences. It depicts the rule relationships between all input
parameters and both TC and TD occurrences. TC and TD will occur when they employ value from 0.51
to 1. For instance, Figure 3 exhibits FL computation by using centroid method for ϑ= -9.5 S-1, u= -16
m/s, SST= 29.5oC, θe=15oC, and q= 39.5 g/kg. With those values, the rules depict the probability of
either TC or TD occurs with a degree of membership approximately at 0.3. The rule illustrated in Figure
3 exhibits the condition of TC and TD to appear is fulfilled. However, this study builds fuzzy rule up to
243 rules that make the computation run complexly.

3.2 The System of Tropical Cyclone (TC) Forecast Design

The Matlab® commercial FL Toolbox was used to design the Graphical User Interface (GUI) of TC
occurrences forecast. This structure is stored inside a GUI tool. All the information for a given fuzzy

6
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

inference system is contained in the GUI design, including variable names, input parameter definitions,
and so on. This system can itself be thought of as supervised learning to predict the occurrences of both
TC and TD.

Figure 4. The GUI main screen of the TC occurrences forecast.

The implementation of the GUI of the TC occurrences forecast is analogous to the GUI used for
Type-1 FL in the Matlab® Fuzzy Logic Toolbox, thus permitting the experienced user to adapt easily
to the use of this GUI. This GUI was built through the process of an algorithm learning from the training
dataset that can be thought of as a teacher supervising the learning process based on the FL rule that we
constructed. The supervised learning problem in this study was classified into two groups, i.e., ‘OCCUR,
and NO OCCUR’. OCCUR is indicate that all the input variables are possible to make TC appeared,
while NO OCCUR is the contrast condition. Figure 4 exhibited the example of all input variables that
enable to make TC occurred in the study area.

3.3 The accuracy of the Forecast

The accuracy of the forecast was calculated from the total numbers of observed and forecast occurrences
and non-occurrences which are given on the lower and right sides of the contingency table. The
validation of the forecast was calculated based on daily data from 2011 to 2017 in DJF and MAM
periods.

Table 3. Forecast validation from contingency table of 2011-2017 periods

Observed
Forecast Total
Yes No
Yes 10 329 339
No 4 1022 1026
Total 14 1351 1365

In table 3, a yes/no forecast of TC occurrences for 21 months are presented, and the high values of
hit and correct negatives indicate that the model may be considered as practically accurate. Interestingly,

7
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

the results reveal that the model predictions are in good agreement with the observed TC and TD events.
Table 3 depicts that this system can predict 1032 event that matched with TC observation and only 333
events that miss predicted during 2011 to 2017 periods. During these periods, 10 TC occurrences can
be predicted accurately from 14 number of TC events. From table 3, the accuracy of the forecast can be
calculated based on the correct number of predictions. The model satisfactorily simulated the occurrence
of TC with comparable error measures. The result exhibits the accuracy at 0.75 (range: 0 to 1, perfect
score: 1).

3.4 Comparing of FL model with previous studies in term of TC and TD forecasting

Most studies of the tropical cyclone (TC) forecasting are focused on the track and the wind radius
forecasting, even though a tropical cyclone formation alerts also developed. For instance, Vitart, Leroy
and Wheeler [24] compared the accuracy of the European Centre for Medium-Range Weather Forecasts
(ECMWF) forecast system to predict the genesis of tropical cyclones (TCs) over the Southern
Hemisphere during intraseasonal (10-60 days) periods have been investigated and compared to the skill
of a statistical approach. Other studies have been developed to investigate the potential link between
TCs and its impacts both at a regional and global scale in worldwide [25-29]. For instance, in terms of
extremes, Khouakhi, Villarini and Vecchi [29] employing annual maximum (AM) and peak-over-
threshold (POT) methods, the highest impact of tropical cyclone to rainfall are found in East Asia,
followed by Australia and North and Central America, with fractional contributions generally decreasing
farther inland from the coast. He found that the relationship between extreme rainfall induced by TC
and ENSO exhibits that extreme rainfall caused by TC tends to occur more frequently in Australia and
along the East Coast of the U.S. during La Niña and in East Asia and the northwest Pacific islands during
El Niño.
For the forecast of TC formation, Li, Fu, Ge, Wang and Peng [30] used QuikSCAT satellite products
and the Tropical Rainfall Measurement Mission Microwave Image data to document temporal and
spatial structures of Rossby wave to investigate cyclogenesis formation in the western North Pacific.
However, there is still a lack of studies to predict the genesis of TC occurrences. Therefore, this paper
takes a different approach and explores and forecast the TC occurrences based on unsupervised machine
learning (self-organized) system. This study presents the development of fuzzy logic (FL) models as an
unsupervised machine learning approach [31] for predicting the occurrence of TC from five primary TC
genesis parameters. This machine learning approach is a family of statistical methods with origins in the
area of artificial intelligence. This method is developed as holding great breakthrough for the
improvement of understanding and prediction about the earth-science system. This modeling approach
is quite flexible to resolve complex problems with various interacting parameters and typically
outcompete traditional methods (e.g., generalized linear models), performing them ideal for modeling
atmospheric environment [32]. In this study, for the results, the skill of the model exhibits good
agreement with the observed TC and TD events.

3.5 Limitations
There are some limitations for this study to develop FL model. A small number of fuzzy sets leads to
unrepresentative forecasts whereas a large number leads to many calculations. In previous studies, many
numbers of fuzzy sets are selected initially from 3 to 6 [33, 34]. Therefore, it is difficult to determine
the length of interval for FL approaches. In this study, we set 3 number of fuzzy and labeled with, low,
medium, and high based on the highest accuracy of the training dataset. In addition, we used 3 number
of fuzzy sets to save computational resources. In FL systems, membership functions play a very
important role in representing problems [35]. The membership function that is used in this study is a
trapezoidal and triangular membership function. We used the trapezoidal and triangular membership
function regarding the behavior of TC that visually matched with its pattern criteria. This TC behavior
can come from expert knowledge and training data. In this study, we construct TC behavior based on
the training data in a daily calculation period from 1989 to 2010. The longer periods that potentially be
used in future research may result in different membership function.
Furthermore, the selection of training data can also be a difficulty, because the data used must
represent the actual data. The selection of training data will determine the knowledge and accuracy of

8
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

the generated fuzzy logic. Ultimately, in this study, the accuracy of fuzzy logic depends on the expert
or training data. In this method, it is also possible to combined FL approach with another method such
as adaptive neuro-fuzzy inference system (ANFIS) to get the better skill of TCs prediction.
From the results, the forecasting of TC occurrences using FL algorithm is generally accurate based
on the accuracy score at 0.75. However, this model only can be used in this study area (15°S-10° S and
115°E-130°E). It is a need either to do a cluster analysis or to explore other TC occurrences at potential
areas such as North Atlantic, West Pacific or the South China Sea before making a forecast. Different
locations are possible to produce different rules of FL algorithm. For further research, it also possible to
employ another input variable to increase the forecast accuracy of TC occurrences.

4. Conclusion
This contribution explored the new forecasting of TC occurrences method utilizing FL algorithm as an
unsupervised machine learning approach. We built-up FL rules and identified a defuzzification
algorithm based on 33th percentile of TC and TD historical data set from 1991-2010. Forecast accuracy
is tested using a different set of training data from the one used to obtain the rule base. The model
satisfactorily simulated the occurrence of TC with comparable error measures in the study area. The
result exhibits the accuracy at 0.75 (range: 0 to 1, perfect score: 1) showing a good agreement with the
observed TC and TD events. In addition, we employed The Matlab® commercial FL Toolbox to design
the Graphical User Interface (GUI) of TC genesis (occurrences) forecast. The evidence shows that the
result provides insights into the adequacy of FL methods for forecasting the TC occurrences.

5. References
[1] Henderson-Sellers A, Zhang H, Berz G, Emanuel K, Gray W, Landsea C, Holland G, Lighthill J, Shieh S-L
and Webster P 1998 Tropical cyclones and global climate change: A post-IPCC assessment Bulletin of
the American Meteorological Society 79 19-38
[2] Neiburger M, Edinger J G and Bonner W D 1995 Memahami Lingkungan Atmosfer Kita Edisi kedua,
Penerbit ITB, Bandung
[3] Anthes R 1982 Tropical Cyclones: Their Evolution, Structure and Effects. American Meteorological Society,
208 pp., and DP Baumhefner, 1984: A diagram depicting forecast skill and predictability Bull. Amer.
Meteor. Soc 65,701-3
[4] Gray W M and Landsea C W 1992 African rainfall as a precursor of hurricane-related destruction on the US
East Coast Bulletin of the American Meteorological Society 73 1352-64
[5] Gray W M, Landsea C W, Mielke Jr P W and Berry K J 1993 Predicting Atlantic basin seasonal tropical
cyclone activity by 1 August Weather and Forecasting 8 73-86
[6] Tonkin H 1996 Investigating thermodynamic models of tropical cyclone intensity. MS thesis, Climatic
Impacts Centre, Macquarie University, Sydney, Australia)
[7] Diaz H F and Pulwarty R S 2012 Hurricanes: climate and socioeconomic impacts: Springer Science &
Business Media)
[8] Management N A f D 2017 Tropical Cyclone Cempaka Hit Indonesia Causing Floods, Landslides and
Tornadoes.
[9] Knaff J A and Zehr R M 2007 Reexamination of tropical cyclone wind–pressure relationships Weather and
Forecasting 22 71-88
[10] Dyahwathi N, Effendy S and Adiningsih E S 2007 Tropical Cyclone Characteristic and Its Impact on Rainfall
Anomaly in Indonesia J. Agromet Indonesia 61-72
[11] Khotimah M 2008 Klimatologi Siklon Tropis di Indonesia Bul. Met Klim Dan Geo 4 251-64
[12] Underground W 2018 Hurricane and Tropical Cyclones.
[13] Gray W M 1975 Tropical cyclone genesis Atmospheric science paper; no. 234
[14] Sinclair M R 1994 An objective cyclone climatology for the Southern Hemisphere Monthly Weather Review
122 2239-56
[15] Reanalysis J 2018 iTacs (Interactive Tool for Analysis of the Climate System).
[16] Bellman R E and Zadeh L A 1970 Decision-making in a fuzzy environment Management science 17 B-141-
B-64
[17] Tayfur G and Singh V P 2006 ANN and fuzzy logic models for simulating event-based rainfall-runoff Journal
of hydraulic engineering 132 1321-30
[18] Alvisi S, Mascellani G, Franchini M and Bardossy A 2006 Water level forecasting through fuzzy logic and
artificial neural network approaches Hydrology and Earth System Sciences Discussions 10 1-17

9
International Conference On Tropical Meteorology And Atmospheric Sciences IOP Publishing
IOP Conf. Series: Earth and Environmental Science 303 (2019) 012044 doi:10.1088/1755-1315/303/1/012044

[19] Xiong L, Shamseldin A Y and O'connor K M 2001 A non-linear combination of the forecasts of rainfall-
runoff models by the first-order Takagi–Sugeno fuzzy system Journal of hydrology 245 196-217
[20] Chang F-J and Chang Y-T 2006 Adaptive neuro-fuzzy inference system for prediction of water level in
reservoir Advances in water resources 29 1-10
[21] Tagaki T and Sugeno M 1985 Fuzzy identification of systems and its application to modelling and control
IEEE Trans. Syst. Man and Cybernetics 15 116-32
[22] Provost F J and Fawcett T 1997 Analysis and visualization of classifier performance: Comparison under
imprecise class and cost distributions. In: KDD, pp 43-8
[23] Ibrahim A 2003 Fuzzy logic for embedded systems applications: Elsevier)
[24] Vitart F, Leroy A and Wheeler M C 2010 A comparison of dynamical and statistical predictions of weekly
tropical cyclone activity in the Southern Hemisphere Monthly Weather Review 138 3671-82
[25] Knight D B and Davis R E 2009 Contribution of tropical cyclones to extreme rainfall events in the
southeastern United States Journal of Geophysical Research: Atmospheres 114
[26] Chang C P, Lei Y, Sui C H, Lin X and Ren F 2012 Tropical cyclone and extreme rainfall trends in East Asian
summer monsoon since mid‐20th century Geophysical Research Letters 39
[27] Lavender S L and Abbs D J 2013 Trends in Australian rainfall: contribution of tropical cyclones and closed
lows Climate dynamics 40 317-26
[28] Villarini G and Denniston R F 2016 Contribution of tropical cyclones to extreme rainfall in Australia
International Journal of Climatology 36 1019-25
[29] Khouakhi A, Villarini G and Vecchi G A 2017 Contribution of tropical cyclones to rainfall at the global scale
Journal of Climate 30 359-72
[30] Li T, Fu B, Ge X, Wang B and Peng M 2003 Satellite data analysis and numerical simulation of tropical
cyclone formation Geophysical research letters 30
[31] Lin C-T and Lee C S G 1991 Neural-network-based fuzzy logic control and decision system IEEE
Transactions on computers 40 1320-36
[32] Olden J D, Lawler J J and Poff N L 2008 Machine learning methods without tears: a primer for ecologists
The Quarterly review of biology 83 171-93
[33] Stuber M, Gemmar P and Greving M 2000 Machine supported development of fuzzy-flood forecast systems.
In: European Conference on Advances in Flood Research: Citeseer) pp 504-15
[34] Kucukali S and Baris K 2010 Turkey’s short-term gross annual electricity demand forecast by fuzzy logic
approach Energy policy 38 2438-45
[35] Zadeh L A 1996 Fuzzy Sets, Fuzzy Logic, And Fuzzy Systems: Selected Papers by Lotfi A Zadeh: World
Scientific) pp 775-82

10