Vous êtes sur la page 1sur 6

Optimization of Neural Network Based on Hybrid Method of

Genetic Algorithm and Particle Swarm Optimization for


Maritime Weather Forecasting in Buoyweather Station Type II

Dvitiya Srestha Prajna Mahistha Magfiroh Fatwaning Ukhti Muhammad Rifki Kurniawan
Engineering Physics Engineering Physics Engineering Physics
Institut Teknologi Sepuluh Nopember Institut Teknologi Sepuluh Nopember Institut Teknologi Sepuluh Nopember
Surabaya, Indonesia Surabaya, Indonesia Surabaya, Indonesia
dvitiya.srestha@gmail.com magfi.ukhti@gmail.com mrifkikurniawan17@gmail.com

Syamsul Arifin Aulia Siti Aisjah


Engineering Physics Engineering Physics
Institut Teknologi Sepuluh Nopember Institut Teknologi Sepuluh Nopember
Surabaya, Indonesia Surabaya, Indonesia
syamp3ai@gmail.com auliasa@ep.its.ac.id

Abstract— This study aims to predict maritime weather costs and Green House Gas (GHG) emissions while still
(wind speed, wind direction, temperature and wave height) maintaining its cargo service level agreement (SLA) [3].
using an artificial intelligence approach. Artificial intelligence Maritime weather predictions in several weather agencies
used is a predictor of hybrid neural networks modified by still use the statistical time series data approach such as
genetic algorithms and particle swarm optimization. Hybrid
predictors work on single output predictors based on
regression model or ARIMA. This approach requires few
differential evolution algorithm optimization. The research historical data. In addition, the results also have a low
method consists of collecting secondary data as learning data, predictive accuracy [4]. The use of this method is more
statistical testing, designing algorithm architecture, and real suitable for forecasting and cannot work on non-linear data
time validation. Data on the learning process of artificial neural and current weather conditions that are volatile and variable
networks was collected from the BMKG Maritime II Perak [5].
Weather Station. Learning data used maritime weather data in This research designs predictors using hybrid method of
the span of 3 years from 2015 to 2017. The best predictors for Artifical Neural Network - Genetic Algorithm (ANN-GA)
each variable are validated using real time data obtained from and Artificial Neural Network - Particle Swarm Optimization
Buoyweather Stationn Type II measurement at the Madura
Strait, Java Sea. The prediction results in the learning and
(ANN-PSO) which were optimized with Differential
validation process indicate that the NN-Hybrid predictor Evolution (DE) algorithm for weather variables including
provides the best results for all variables. However, realtime temperature, speed and wind direction, and wave height.
validation shows different results, where hybrid predictors do These predictors work through crosses between single
not always have the best performance -the smallest RMSE- predictors. Crossing occurs through the provision of
compared to the single predictor. coefficients in each single predictor, where coefficient is
determined based on optimization using Differential
Keywords— artificial neural network, genetic algorithm, Evolution algorithm.
particle swarm optimization, differential evolution, maritime
The use of hybrid method can take advantage of each
weather prediction.
methodology used [6]. The idea of using hybrid methods was
I. INTRODUCTION initiated firstly by J.M. Bates and C.W.J. Granger [7] which
proved that combination methods are more efficient and
Maritime weather information both now and in the future
easier than a single method. In addition, Armstrong's meta
is an important needs in this decade. This Information can be
analysis showed that combining various techniques was more
used as a consideration in making decisions in various
useful for forecasting short-term problems [8].
maritime activities. This is useful for fishermen to determine
The methodology used is of Artificial Neural Networks
the feasibility of the shipping lane in determining the safe
optimized using Particle Swarm Optimization and Genetic
point in safety aspects. Weather information such as wave
Algorithms. Particle Swarm Optimization (PSO) is an
height is a variable that is useful for the assessment of safety
evolutionary algorithm which is an important tool for the
in maritime activities in addition to other considerations, such
optimization and resolution of complex problems. PSO is an
as the characteristics of the ship [1], [2]. Maritime safety
interesting computational technique with high capability and
assessment also be used as consideration in maritime
practical application because it has simple and easy to use in
industrial activities such as managing up and down landing,
various fields.. Meanwhile, ANN combined with the
dredging involving floating objects, or renewable energy
optimization of Genetic Algorithm (GA) is able to predict
generator. Thus, the chances of failure due to risks can be
better than ordinary ANN. The results of ANN-Genetic
reduced and higher efficiency can be achieved. In economi
Algorithm optimization combination are able to obtain more
aspects, maritime weather data is used as a decision support
accurate prediction accuracy with faster convergence [9].
system for vessel speed optimization. While in the logistics
Optimization with DE algorithm is used as a methodology
sector, it can be used in optimizing ship speed to reduce fuel
in determining the weights coefficient on hybrid predictor.
DE has the special ability to memorize knowledge to keep

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE


Data ready for use after statistical tests, then divided into
TABLE 1. r correlation coefficient TABLE 2. r correlation coefficient 80% training data and 20% data validation for predictor
of wind speed variable. of wind direction variable. algorithm training process.
Independent Wind speed Wind
Variable r coefficient Independent direction
B. Correlation Analysis
Wind speed Variable r Quantitative modeling to find out how big the relationship
0,885 coefficient
(1 hour ago) between two variables between independent and dependent
Wind direction (1
Wind speed hour ago)
0,858 variables is done on the correlation. Determination of the
0,771
(2 hour ago) Wind direction (2 coefficient value at the relationship level used in the
0,716
Wind speed
hour ago) correlation refers to the reference. Some independent
0,669 Wind direction (3 variables in this study have very strong and strong varied
(3 hour ago) 0,609
hour ago)
relationships with each dependent variable.
The independent variables chosen are the variable values
TABLE 3. r correlation TABLE 4. r correlation coefficient
coefficient of temperature of wave height variable.
at 1, 2, and 3 hours ago. This research does not take up to 4
variable. hours and more, because the more time past the weaker the
coefficient value r. This statement is supported by the results
Temperatur Independent Wave height
Independen
er Variable r coefficient of the correlation test in this study. Correlation test results can
t Variable Wind wave be seen in TABLE 1 till TABLE 4. The relationships between
coefficient
Wind height (1 hour 0,999 variable values at 1, 2, and 3 hours ago show a large r
temperature 0,857 ago) coefficient so that it can be used as independent variable
(1 hour ago) Wind wave input.
Wind height (2 hour 0,999
temperature 0,734 ago)
(2 hour ago)
Wind Wind wave III. OPTIMIZED LEARNING METHOD OF ARTIFICAL NEURAL
temperature 0,593 height (3 hour 0,997 NETWORK
(3 hour ago) ago)
The methodology used as a prediction is backpropagation
track of dynamically [6]. DE also needs fewer parameters
artificial neural network. The design of artificial neural
than GA and PSO. This process is carried out in order to
network algorithms is divided into 3 types of methodology,
adjust the proportion of coefficients to get the best
those are artificial neural networks optimized with genetic
performance results.
method algorithm (ANN-GA), artificial neural networks
optimized by particle swarm optimization (ANN-PSO), and
II. STATISTICAL TEST AND PREPROCESSING DATA hybrid method which is the combination of ANN-GA and
ANN-PSO optimized using the Differential Evolution (DE).
A. Collecting Data Artificial neural network architecture consists of 3 layers:
The research data used are secondary time series data input layer, hidden layer, and output layer. The input layer
obtained from the Meteorology, Climatology and Geophysics consists of 3 inputs for each variable respectively 1, 2, and 3
Agency (BMKG) Maritime II Perak. Data collected in the hours ago. This is determined based on the results of the
form of hourly data in the period of 2011 - 2017. Hourly data correlation test that has been carried out. This is a test that has
is collected in the period 2011 to 2017. The predicted weather been done. The hidden layer of each network using hidden
variables are air temperature, wind speed, wind direction, and nodes varies from 1,3,5,7, and 9. While the output layer is the
wave height. Secondary data that has been collected is output of each predicted variable.
evaluated using statistical testing. Statistical test of the data
carried out are missing data, outlier, and correlation test.
Missing data is the loss of certain data in a learning dataset.
This problem was solved by autoregressive imputation
(ARLSImpute) which is suitable for time series dataset [10].
Define AR coefficients is main purpose of this method with
several steps of calculations such as defining K-similar data,
calculating AR coefficients, and finally estimating missing
values. The iteration process continues until the convergence
is reached.

Outlier data is data that deviates from other data values,


causing data to be biased. Outlier data can occur due to errors
in data retrieval or there are extreme data that cannot be
avoided [11].
Correlation test is to determine the relationship between Fig 1. Neural network architecture
variables linearly and expressed by the correlation coefficient
(r) [12]. The value of the correlation coefficient is more In the following Fig 1 can be described network
closely related if it approaches the value of +1 or -1, where architecture for air temperature variables with 3 hidden
each describes a direct or opposite relationship between nodes. The X notation is network input unit where X1 shows
variables. the value at t-1 hours, X2 at t-2 hours, and X3 at t-3 hour. Z
is hidden unit, and Y is the output at time t. The weight of the optimization. It works based on a population that runs
input unit to the hidden unit is indicated by the notation v and towards mutations, crossover, and selection processes. The
the weight of the hidden unit to the output unit indicated by RMSE performance parameter is used to determine the
the notation w. population that is applied to the hybrid equation.
The algorithm design of JST-GA uses artificial neural
network as the main algorithm and optimization of Genetic
Algorithm as learning methodology to determine the optimal
weight used by network model. All the weights value on the
network are given as individuals that will form the
population. Then optimization works by mutation and
crossover to determine the best individual / optimal weight.
The fitness value is given by the RMSE value. Where the
RMSE value is getting better when the value is getting closer
to zero.

Fig 4. DE optimization algorithm as coefficient determiner for hybrid


equation.

Population initialization process to create population that


consist of 250 individuals, NP=250 with problem dimension
of 2, D=2, can be formulated as.

{X 1,i,0 =(x 1,i,0 , x 2,i,0 ) |i=1,2,3,…,250} (1)

𝑢𝑢𝑢𝑢
with 𝑥𝑥𝑙𝑙𝑙𝑙𝑙𝑙
𝑗𝑗 ≤ 𝑥𝑥𝑗𝑗,𝑖𝑖,0 ≤ 𝑥𝑥𝑗𝑗 for j=1,2
Fig 2. Flow diagram of Genetic Algorithm optimization to determine AAN
weights.
𝑥𝑥𝑙𝑙𝑙𝑙𝑙𝑙
𝑗𝑗 =0
The ANN-PSO algorithm design is similar to ANN-GA
𝑢𝑢𝑢𝑢
which determines the optimal weight value to get the best 𝑥𝑥𝑗𝑗 = 1
results. Weights as particles in the formed population. The
best global values and the best particles are used as the best Where,
weights. The value of fitness is given by the RMSE value. NP= population number
D= problem dimension
𝑥𝑥𝑙𝑙𝑙𝑙𝑙𝑙
𝑗𝑗 = lower coefficient constraint
𝑢𝑢𝑢𝑢
𝑥𝑥𝑗𝑗 = upper coefficient constraint

The mutation process is used to get the new population


obtained from the equation below. In this study, mutation
factors were defined by mf = 0.75. And DE / best / 1 is based
on the formula of this DE algorithm..

DE/best/1 𝑣𝑣𝑖𝑖,𝑔𝑔 = 𝑋𝑋𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏,𝑔𝑔 + 𝑚𝑚𝑚𝑚𝑖𝑖 (𝑋𝑋𝑟𝑟1,𝑔𝑔 − 𝑋𝑋𝑟𝑟2,𝑔𝑔 ) (2)

Where
Fig 3. Flow diagram of PSO optimization which is has results as ANN r1, r2= integers chosen randomly from the set of numbers
weights.
{1,2,3,…,NP}\{i}
Hybrid algorithm design is a combination of ANN-GA and 𝑋𝑋𝑟𝑟1,𝑔𝑔 − 𝑋𝑋𝑟𝑟2,𝑔𝑔 = difference between parent population
ANN-PSO output. Each single network is used as an input 𝑋𝑋𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏,𝑔𝑔 = best population based on RMSE
unit. The best weight value is obtained from DE optimization, 𝑚𝑚𝑚𝑚= mutation factor (mf=0.75)
where the algorithm is similar to genetic algorithm
16
14
12
Wind Speed (knot)

10
NN-GA
8
NN-PSO
6
NN-Hybrid
4
Target
2
0
1 6 11 16 21 26 31 36 41 46
Day

Fig 5. Wind speed data test

400
350
Wind Direction (degree)

300
250
NN-GA
200
NN-PSO
150
NN-Hybrid
100
50 Target

0
0 5 10 15 20 25 30 35 40 45 50
Day

Fig 6. Wind direction data test

At the crossover stage, crossing and exchanging


chromosomes in the population is carried out. This exchange IV. PROPOSED HYBRID PREDICTOR
occurs if fulfilling the condition with certain parameters used
by CR, crossover ratio as a parameter, in this case is
determined the value of CR = 0.1. Mathematically formulated
as follows
Fig 7. Block diagram of proposed hybrid predictor.
𝑈𝑈𝑖𝑖,𝑔𝑔 = (𝑢𝑢1,𝑖𝑖,𝑔𝑔 , 𝑢𝑢2,𝑖𝑖,𝑔𝑔 ) (3)
𝑌𝑌�ℎ𝑦𝑦𝑦𝑦𝑦𝑦𝑦𝑦𝑦𝑦 = 𝑤𝑤1. 𝑌𝑌�𝑁𝑁𝑁𝑁−𝐺𝐺𝐺𝐺 + 𝑤𝑤2. 𝑌𝑌�𝑁𝑁𝑁𝑁−𝑃𝑃𝑃𝑃𝑃𝑃 (6)
𝑣𝑣𝑗𝑗,𝑖𝑖,𝑔𝑔 𝑗𝑗𝑗𝑗𝑗𝑗𝑗𝑗 (𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟(0,1) ≤ 𝐶𝐶𝐶𝐶 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗 = 𝑗𝑗𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟
𝑢𝑢𝑗𝑗,𝑖𝑖,𝑔𝑔 = � (4)
𝑥𝑥𝑗𝑗,𝑖𝑖,𝑔𝑔 , 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠
With ∑1𝑖𝑖=1 𝑤𝑤𝑖𝑖 = 1, 0 ≤ 𝑤𝑤𝑖𝑖 ≤ 1
From the binomial operation above, the crossover process is Where:
carried out on the population chromosome that has a mutation Ŷ hybrid = predictor output of hybrid by DE algorithm
if it meets the conditions j = jrand or randj (0.1).
Ŷ ANN-GA = predictor output of ANN-GA
The last stage in the DE algorithm is selection. The selection Ŷ ANN-PSO = predictor output of ANN-PSO
process carried out a selection of new populations between
the parent population and the mutation and crossover Realtime testing used weather data obtained from
population, this selection is based on the Ui and Xi function buoyweather station type II measurement. Realtime data
which refers to the RMSE function. Thus the individual measurement was carried out at the Madura Strait for three
chosen was a population with the smallest RMSE value.. days.
The training, validation, and realtime validation process
𝑈𝑈𝑖𝑖 , 𝑔𝑔, 𝑗𝑗𝑗𝑗𝑗𝑗𝑗𝑗 𝑓𝑓(𝑈𝑈𝑖𝑖 ) ≤ 𝑓𝑓(𝑋𝑋𝑖𝑖 ) provide the best network results that can be used based on the
𝑋𝑋𝑖𝑖 = � (5)
𝑋𝑋𝑖𝑖 , 𝑔𝑔, 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 RMSE values of each ANN-GA, ANN-PSO, and ANN-
Hybrid performance.
35
34
33
Temperature (oC)

32
31 NN-GA
30 NN-PSO
29 NN-Hybrid
28
Target
27
26
1 6 11 16 21 26 31 36 41 46
Day

Fig 8. Temperature data test

1
0.9
0.8
Wave Height (hsig)

0.7
0.6 NN-GA
0.5
0.4 NN-PSO
0.3 ANN-Hybrid
0.2
Target
0.1
0
1 6 11 16 21 26 31 36 41 46
Day

Fig 9. Wave height data test

V. PREDICTION RESULTS AND DISCUSSION updating particle speed and position. The solution obtained is
only through one-way information from the best particles to
A. Comparison of Each Predictor Performance Analysis other particles so that the solution obtained is local optimal
RMSE value of training , validation , and realtime data [14]. Each optimization deficiency is compensated by
validation using buoyweather can be seen in TABLE 5. The combining GA and PSO optimization or what is called ANN-
RMSE value of each predictor is presented in each variable Hybrid [15].
of wind speed, wind direction, temperature, and wave height GA has advantages when the initial population is large,
respectively in columns 3, 4, 5, and 6. Look at the RMSE whereas PSO is the opposite, a large population causes
value in the training process on each variable in the table. It performance to decrease [16]. This study tries to observe the
can be seen that in the wind speed variable, the smallest RMSE generated from the same number of population and
RMSE is obtained by ANN-GA and ANN-Hybrid predictors. iterations on ANN-GA and ANN-PSO predictors. The results
The RMSE has the same value. In other variables, wind show that in large population condition, ANN-GA is superior
direction, air temperature, and wave height have the same to ANN-PSO. Where the ANN-GA and ANN-Hybrid
RMSE values on ANN-GA and ANN-Hybrid predictors. We predictors have the same RMSE value. This is because the
refer to the previous references entitled "on the hybrid of ANN-GA has reached a global optimum, so that when GA
genetic algorithms and particle swarm optimization for and PSO are combined, the RMSE does not go down again.
envolving recurrent neural networks", it shows that ANN- Now look at the RMSE value in the validation process.
Hybrid has the smallest RMSE compared to ANN-GA and Similar things also happen to each variable. ANN-GA and
ANN-PSO [13]. ANN-Hybrid have better results on RMSE. Both predictors
Similar results are also shown by our research which also have the same value. This same result is because the
shows that ANN-Hybrid has the best performance. There are training data for each predictor is very large so that the results
several reasons why ANN-Hybrid has produced the best of RMSE training and validation are not much different.
performance. Each optimization has weaknesses, such as Predictors are tested with real time measurements. RMSE
finding the best solution with GA optimization has a long and prediction values can be seen in TABLE 5. The difference in
complex convergence level because it goes through many RMSE training and previous validation with validation with
processes (mutation, crossing, and elitsm), while finding real time data is still quite large. The first variable is wind
solutions with PSO optimization is not as complex as GA. speed, the difference in RMSE validation using measurement
The process of finding the best solution is obtained by data compared to the previous validation is almost three times
smaller in each predictor. Similar things are also produced by REFERENCES
wind direction variables. The difference is almost three times [1] J. F. Balmat, F. Lafont, R. Maifret, and N. Pessel, “A
its value in ANN-GA predictors, while the predictors of decision-making system to maritime risk assessment,”
ANN-PSO and ANN-Hybrid have doubled. Large Ocean Eng., vol. 38, no. 1, pp. 171–176, 2011.
differences are also experienced by wave heights. The [2] P. De Girolamo, M. Di Risio, G. M. Beltrami, G.
difference in the ANN-GA and ANN-Hybrid predictors Bellotti, and D. Pasquali, “The use of wave forecasts for
reached more than ten times, while the ANN-PSO predictor maritime activities safety assessment,” Appl. Ocean
reached more than five times. Different results are indicated Res., vol. 62, pp. 18–26, 2017.
by temperature variables. The difference is quite small and [3] H. Lee, N. Aydin, Y. Choi, S. Lekhavat, and Z. Irani,
not even doubled. The author tries to analyze what causes “A decision support system for vessel speed decision in
RMSE with measurement data to have RMSE greater than the maritime logistics using weather archive big data,”
previous validation. Comput. Oper. Res., vol. 98, pp. 330–342, 2018.
Note Fig 5 on the variable wind speed. Wind speed is a [4] A. S. Aisjah, “Maritime Weather Prediction using
natural phenomenon that has significant changes. This can be Fuzzy Logic in Java Sea,” no. November, pp. 205–208,
seen in the target on the 44th to 48th days which had a drastic 2011.
change, which caused the RMSE to be bigger than training [5] C. Voyant et al., “Machine learning methods for solar
and previous validation. In other variable, the wind direction radiation forecasting: A review,” Renew. Energy, vol.
in Fig 6 is also a natural phenomenon which also has 105, pp. 569–582, 2017.
significant changes. Target data on the 6th to 7th day has a [6] Y. Yang, Y. Chen, Y. Wang, C. Li, and L. Li,
significant change, also on the 14th to 20th and 27th to 29th days “Modelling a combined method based on ANFIS and
also showed a big change. In contrast to the temperature neural network improved by DE algorithm: A case
variable in Fig 8, the changes are not too large so that the study for short-term electricity demand forecasting,”
RMSE is not much different from the training and previous Appl. Soft Comput. J., vol. 49, pp. 663–675, 2016.
validation. Different things occur at wave heights in Fig 9. [7] J. M. Bates and C. W. J. Granger, “The Combination of
Real time data shows changes in ups and downs in the near Forecasts,” or, vol. 20, no. 4, p. 451, 1969.
future, while training data gives a change in ups and downs [8] J. S. Armstrong, “Combining Forecasts : The End of the
in a long time. Beginning or the Beginning of the End ? Combining
Forecasts : The End of the Beginning or the Beginning
TABLE 5. RMSE pelatihan, validasi, dan buoyweather
of the,” vol. 5, no. 1989, pp. 585–588, 1989.
RMSE [9] S. Zhang, H. Wang, L. Liu, C. Du, and J. Lu,
Pro Predictor Wind
Temp
Wavehei
“Optimization of Neural Network Based on Improved
Wing Speed eratu Genetic Algorithm,” 2009 Int. Conf. Comput. Intell.
cess Types Direction ght
(knot) re
(degree)
(oC)
(hsig) Softw. Eng., no. Cctot, pp. 203–207, 2009.
ANN-GA 0,97 39 1,00 0,01 [10] I. Pratama, A. E. Permanasari, I. Ardiyanto, and R.
ANN- Indrayani, “A review of missing values handling
Trai 1,03 41 1,02 0,02
ning
PSO methods on time-series data,” 2016 Int. Conf. Inf.
ANN- Technol. Syst. Innov. ICITSI 2016 - Proc., 2017.
0,97 39 1,00 0,01
Hybrid
ANN-GA 0,66 51 1,07 0,01 [11] S. Santoso, Statistika Multivariat, Revisi. Jakarta,
Vali ANN- Indonesia: PT Elex Media Komputindo, 2010.
0,69 52 1,10 0,02
datio PSO [12] B. Kamranzad, A. Etemad-Shahidi, and M. H.
n ANN- Kazeminezhad, “Wave height forecasting in Dayyer,
0,66 51 1,07 0,01
Hybrid
the Persian Gulf,” Ocean Eng., vol. 38, no. 1, pp. 248–
ANN-GA 2,36 140 1,11 0,13
ANN- 255, 2011.
2,36 111 1,13 0,12 [13] C.-F. Juang and Y. C. Liou, “On The Hybrid of Genetic
Test PSO
ANN-
2,36 107 1,11 0,15
Algorithm and Particle Swarm Optimization for
Hybrid Envolving Recurrent Neural Network,” pp. 2285–2289,
2004.
[14] K. O. Jones, “Comparison of Genetic Algorithm and
VI. CONCLUSION
Particle Swarm Optimisation,” Int. Conf. Comput. Syst.
The ANN-Hybrid combination result shows that these Technol. - CompSysTech’2005 Comp., pp. 1–6, 2005.
predictors have been able to predict the air temperature in the [15] J. Wang, F. Zhang, F. Liu, and J. Ma, “Hybrid
next few hours well, while other variables such as wind speed, forecasting model-based data mining and genetic
wind direction, and wave height cannot be said to be algorithm-adaptive particle swarm optimisation: a case
predictable using this ANN-GA, ANN-PSO, and ANN- study of wind speed time series,” IET Renew. Power
Hybrid design. Gener., vol. 10, no. 3, pp. 287–298, 2016.
[16] Sharma, Jyoti and R. S. Singhal, “Comparative
Research on Genetic Algorithm, Particle Swarm
ACKNOWLEDGMENT
Optimization and Hybrid GA-PSO,” Int. Conf. Comput.
This research was assisted by several institutions. For that Sustain. Glob. Dev., vol. 2, pp. 110–114, 2015.
reason, thanks to BMKG Perak I and II Surabaya as data
supply, Kemerinstek Dikti which has funded this research
through the superior applied research fund of college in 2018.

Vous aimerez peut-être aussi