number of epochs when the data cluster is 100%. The result indicates that when the
data cluster is 30 - 60%, the stability condition still fluctuates and has not yet reached
convergence. After reaching 100% clu ster data, the number of epochs decreases and
decreases the epoch time, indicating that Bayesian regularization achieved the best
convergence under conditions of minimum time duration.
The table also shows that the accuracy of each data cluster fluctuates from low -
up - down in the MSE interval of 10
-3
- 10
-4
. After MSE 10
-5
, the accuracy fluctuation
became stable from low - medium - up, meaning that the mo del achieved with the MSE
10
-5
target has a stable accuracy. This is also marked by the difference in the larg est
data cluster, with previously having the smallest value of 0.239 compared to the other
two values of - 2.559 and - 0.967.
Table 4: Performance at Rainy Season
Weather
Performance at Rainy Season (%)
MSE < 10
-3
MSE < 10
-4
MSE < 10
-5
30% 60% 100% 30% 60% 100% 30% 60% 100%
Max Epoch 220 499 196 286 887 578 724 1027 548
Best Epoch 219 499 195 262 886 577 723 1026 548
T ime (s) 5.294 9.5 6.299 5.634 13.558 12.516 10.933 15.659 12.030
Best Accuracy
98.642
*2)
100.00
*2)
100.00
*3)
98.812
*2)
99.600
*2)
99.260
*2)
98.812
*2)
100.000
*2)
100.00
*2)
Mean accuracy (%) 96.906 99.845 97.286 96.210 98.879 97.912 95.360 99.459 99.698
Note: *1 ) = Sun n y, *2 )= Clo ud y, *3 )= L ig h t Ra in , *4 ) = He a vy Ra in
Table 5: Performance at Dry Season
Weather
Performance at Dry Season (%)
MSE < 10
-3
MSE < 10
-4
MSE < 10
-5
30% 60% 100% 30% 60% 100% 30% 60% 100%
Max Epoch 655 164 519 3369 279 2459 184 381 359
Best Epoch 654 151 427 3368 225 2458 183 379 341
T ime (s) 9.944 5.782 13.112 34.002 7.225 41.232 4.998 10.146 10.159
Best Accuracy
99.267
*1)
100.00
*2)
99.901
*1)
99.853
*1)
99.573
*2)
100.00
*3)
99.120
*1)
100.00
*3)
99.88
*2)
Mean accuracy (%) 98.390 98.542 99.402 98.696 97.699 98.688 98.466 99.233 99.056
Note: *1 ) = Sun n y, *2 )= Clo ud y, *3 )= L ig h t Ra in , *4 ) = He a vy Ra in
Table 5 shows that the best fluctuations in the number of epochs are in the MSE
10
-5
interval. The data will increase and then decrease steadily so that the time
duration will also be shorter. In addition, the difference in mean Accuracy between
MSE clusters has the smallest difference in MSE 10
-5
worth - 0.177 compared to the
difference i n the other two MSEs, respectively 0.86 (MSE 10
-3
) and 0.98 (MSE 10
-4
).
Table 6 : Perf orm ance com para tion of Backpr opag ation a nd Bayesian reg ularizati on
Performance
Error dan T ime
Winter Summer
BP BR BP BR
MSE 0.264606 0.00331 0.12484 0.01509
time (s) 186,9 12,030 152,154 10,146
The results obtaine d fr om t able - 6 show that BR has th e smallest MSE to obtai n
the best accu racy th an usin g BP [34] . Bayesian R egula rizati on has an MSE value o f
0.0033 1 with 12.0 30 sec onds in t he r ainy se ason; MSE 0.0 1509 wit h a time of 1 0.14 6
seconds in the d ry season. Co mpari ng these results wit h Backpropag ation which
has an MSE of 0.2640 6 with 186.9 seco nds in the rai ny season; MSE 0.124 8 4 with
a time of 152. 154 seco nds in the dry se ason. This sig nificant value pr oves that
Bayesian Regulariz ation has th e best MSE and much bett er spee d than
Backpropag ation.
M o hsin [23 ] also prov ed tha t Bayesian Neu ral Network has bette r spee d than
using L east Abs olute Sh rinkag e a nd Selecti on O perat or ( LASSO). Purba [22 ]
compar ed th at th e Bayesian R egulariz ation s peed of a f e w neuro ns (typ e - 2 ) had the
best comp utatio nal time c ompa red t o that of a n euro n with a more si gnificant numb er
(type - 3 ), alt houg h the best acc uracy was obtain ed fo r a la rge nu mbe r of n euro ns.
Our st udy has a bette r sig nificance with the smalles t MS E val ue of 0.00 33 than
Z hao resea rch [2 4] only 0.05. Sev eral p oints r esulted fro m this res earch , includin g:
a. Based on the figures and tables results, adding the number of datasets will affect
the system accuracy level and reduce the mean square error.
b. Increasing the number of datasets will reduce the system speed in providing
forecasting output because the syste m will require the epoch time also longer.
c. The graph shows that Bayesian regularization reduces overfitting, which
increases the system accuracy, which is directly proportional to the decline in
MSE.
6. Conclusion
Weather for ecasting feat ures of air pressu re, humidity , cloud h ood, wind s peed, an d
precipitatio n using a Bayesian re gularizati on neu ral netwo rk have bette r perfo rma nce than
backpro pagati on. They indicate that the use of t his m et hod p rovides high accur acy a nd
good c onve rgence spe ed. Fo r th e su bsequ ent develop me nt, th e m odel will i mplem ent deep
learning recu rren t for forec asting t o com pare with th e r esults of p revious s tudies. The
numbe r of d atasets must inc reas e the best p erfo rmanc e.
Acknowledgment
This study w as cond ucted unde r the U nivers itas Dip a Ma kassar , In donesi a.
References
[1] S. E. Haupt, et al. ³ 0 D F K L Q H O H D U Q L Q J I R U D S S O L H G Z H D W K H U S U H G L F W L R Q ´ L Q 3 U R F H H G L Q J V - IEEE
14th International Conference on eScience, e - Science 2018, pp. 276 ± 277, 2018.
[2] R. Herwanto, R. P. - I. C. on, and undefin H G ³ 5 $ , 1 ) $ / / 3 5 ( ' , & 7 , 2 1 8 6 , 1 * ' $ 7 $
0 , 1 , 1 * 7 ( & + 1 , 4 8 ( 6 ´ M X U Q D O G D U P D M D \ D D F L G $ F F H V V H G - X O
[3] . 8 - D V H H Q D D Q G % & . R Y R R U ³ ' H W H U P L Q L V W L F Z H D W K H U I R U H F D V W L Q J P R G H O V E D V H G R Q L Q W H O O L J H Q W
S U H G L F W R U V $ V X U Y H \ ´ - R X U Q D O R I . L Q J 6 D X G University - Computer and Information Sciences.
King Saud bin Abdulaziz University, Sep. 24, 2020 .
[4] M. Holmstrom, D. Liu, C. V. - 0 $ S S O D Q G X Q G H I L Q H G ³ 0 D F K L Q H O H D U Q L Q J D S S O L H G W R Z H D W K H U
I R U H F D V W L Q J ´ F V V W D Q I R U G H G X $ F F H V V H G - X O
[5] A. * 6 D O P D Q % . D Q L J R U R D Q G < + H U \ D G L ³ : H D W K H U I R U H F D V W L Q J X V L Q J G H H S O H D U Q L Q J
W H F K Q L T X H V ´ L Q , & $ & 6 , 6 - 2015 International Conference o n Advanced Computer Science
and Information Systems, Proceedings, pp. 281 ± 285, 2016.
[6] 0 & K H Q H W D O ³ $ U H L Q I R U F H ment learning approach to irrigation decision - making for rice using
Z H D W K H U I R U H F D V W V ´ $ J U L F : D W H U 0 D Q D J Y R O S 0 D \ .
[7] 7 - X Q J D Q G 0 0 D W V X H G D ³ 9 H U L I L F D W L R Q R I J O R E D O Q X P H U L F D O Z H D W K H U I R U H F D V W L Q J V \ V W H P V L Q S R O D U
regions using TIGGE G D W D ´ 4 - 5 0 H W H R U R O 6 R F Y R O Q R S S ± 582, Jan. 2016 .
[8] % 6 F K X O ] 0 ( O $ \ D U L 6 / H U F K D Q G 6 % D U D Q ³ 3 R V W - processing numerical weather prediction
H Q V H P E O H V I R U S U R E D E L O L V W L F V R O D U L U U D G L D Q F H I R U H F D V W L Q J ´ 6 R O ( Q H U J \ Y R O S S 016 ± 1031,
May 2021 .
[9] - 6 F K D O N Z L M N + - - - R Q N H U $ 3 6 L H E H V P D D Q G ( 9 D Q 0 H L M J D D U G ³ : H D W K H U ) R U H F D V W L Q J
Using GPU - Based Large - ( G G \ 6 L P X O D W L R Q V ´ % X O O $ P 0 H W H R U R O 6 R F Y R O Q R S S ±
723, May 2015.
[10] 6 . L P - / H H D Q G 7 < R R Q ³ 5 R D G V X U I ace conditions forecasting in rainy weather using artificial
Q H X U D O Q H W Z R U N V ´ 6 D I 6 F L Y R O $ X J
[11] 0 0 D W R X T H W D O ³ 7 K H F O L P D W H F K D Q J H L P S O L F D W L R Q R Q - R U G D Q $ F D V H V W X G \ X V L Q J * , 6 D Q G
$ U W L I L F L D O 1 H X U D O 1 H W Z R U N V I R U Z H D W K H U I R U H F D V W L Q J ´ - 7 aibah Uni v. Sci., vol. 7, no. 2, pp. 44 ± 55,
Apr. 2013.
[12] ) ' D O L S L D Q G 6 < < D \ L O J D Q ³ 7 K H , P S D F W R I ( Q Y L U R Q P H Q W D O ) D F W R U V W R 6 N L L Q J , Q M X U L H V % D \ H V L D Q
Regularization Neural Network Model for Predicting Skiing Inj uries SEMCON: Semantic and
contextual object ive metric View project AEP (Academic Excha nge for Progress) View project
The Impact of Environmental Factors to Skiing Injuries: Ba yesian Regularization Neural
1 H W Z R U N 0 R G H O I R U 3 U H G L F W L Q J 6 N L L Q J , Q M X U L H V ´
[13] 0 1 D U Y H N D U D Q G 3 ) D U J R V H ³ ' D L O \ : H D W K H U ) R U H F D V W L Q J X V L Q J $ U W L I L F L D O 1 H X U D O 1 H W Z R U N ´
[14] ' 1 ) H Q W H D Q G ' . X P D U 6 L Q J K ³ : H D W K H U ) R U H F D V W L Q J 8 V L Q J $ U W L I L F L D O 1 H X U D O 1 H W Z R U N ´ L Q
Proceedings of the International Conference on Inventive Com munication and Computational
Technologies, ICICCT 201 8, pp. 1757 ± 1761, 2018 .
[15] - 9 D Q G H J U L I I . : D J V W D I I * + R D Q G - 3 O D X J H U ³ ) R U H F D V W L Q J V S D F H Z H D W K H U 3 U H G L F W L Q J
L Q W H U S O D Q H W D U \ V K R F N V X V L Q J Q H X U D O Q H W Z R U N V ´ $ G Y 6 S 5 H V Y R O Q R S S ± 2327,
2005.
[16] F. Rodríguez, A. Fleetwood, A. * D O D U ] D D Q G / ) R Q W i Q ³ 3 U H G L F W L Q J V R O D U H Q H U J \ J H Q H U D W L R Q
W K U R X J K D U W L I L F L D O Q H X U D O Q H W Z R U N V X V L Q J Z H D W K H U I R U H F D V W V I R U P L F U R J U L G F R Q W U R O ´ 5 H Q H Z ( Q H U J \
vol. 126, pp. 855 ± 864, Oct. 2018.
[17] . $ E K L V K H N 0 3 6 L Q J K 6 * K R V K D Q G $ $ Q D Q G ³ : H D W K H U ) R U H F asting Model using Artificial
1 H X U D O 1 H W Z R U N ´ 3 U R F H G L D 7 H F K Q R O Y R O S S ± 318, Jan. 2 012.
[18] 0 ) D W K L 0 R V W D I D + . D V K D Q L 6 H \ H G 0 - D P H L L D Q G ( E U D K L P 0 D K G L S R X U ³ % L J ' D W D
$ Q D O \ W L F V L Q : H D W K H U ) R U H F D V W L Q J $ 6 \ V W H P D W L F 5 H Y L H Z ´ $ U F K & R P S X W Methods Eng., vol. 1,
p. 3.
[19] ( 5 5 R G U L J X H V , 2 O L Y H L U D 5 / ) & X Q K D D Q G 0 $ 6 1 H W W R ³ ' H H S ' R Z Q V F D O H D G H H S
learning strategy for high - U H V R O X W L R Q Z H D W K H U I R U H F D V W ´ $ F F H V V H G - X O
[20] 3 + H Z D J H H W D O ³ 7 H P S R U D O F R Q Y R O X W L R Q D O Q H X U D O 7 & 1 Q etwork for an effective weather
forecasting using time - V H U L H V G D W D I U R P W K H O R F D O Z H D W K H U V W D W L R Q ´ 6 R I W & R P S X W Y R O Q R
pp. 16453 ± 16482, Nov. 2020.
[21] N. Zuraidin Mohd - 6 D I D U ' 1 G ] L , . D J D O L G L V < < D Q J D Q G $ = D N D U L D ³ 6 K R U W - Term Localized
Wea W K H U ) R U H F D V W L Q J % \ 8 V L Q J ' L I I H U H Q W $ U W L I L F L D O 1 H X U D O 1 H W Z R U N $ O J R U L W K P , Q 7 U R S L F D O & O L P D W H ´
Accessed: Jul. 02, 2021.
[22] ' ( O 5 H ] H Q 3 X U E D + 0 D Z H Q J N D Q J D Q G 7 X O X V ³ ) R U H F D V W L Q J 7 L P H 6 H U L H V ' D W D Z L W K $ U W L I L F L D O
1 H X U D O 1 H W Z R U N R I % D \ H V L D Q 5 H J X O D U L ] D W L R Q ´ vol. 15, no. 1, pp. 564 ± 568, 2021.
[23] 6 0 R K V L Q 6 1 5 D P O L D Q G 0 , P G D G ³ 0 H G L X P - Term Wind Sp eed Prediction using Bayesian
1 H X U D O 1 H W Z R U N % 1 1 ´ , Q W - 6 \ V W , Q Q R Y Y R O Q R S S ± 20, 2021 .
[24] + = K D R / 5 6 K L D Q G + - : D Q J ³ 6 X Q O L J K W * U H H Q K R X V H Temperature Prediction Model Based
R Q % D \ H V L D Q 5 H J X O D U L ] D W L R Q % 3 1 H X U D O 1 H W Z R U N ´ $ S S O 0 H F K 0 D W H U Y R O S S ± 874,
2015.
[25] 3 + H Z D J H 0 7 U R Y D W L ( 3 H U H L U D D Q G $ % H K H U D ³ ' H H S O H D U Q L Q J - based effective fine - grained
Z H D W K H U I R U H F D V W L Q J P R G H O ´ 3 D ttern Anal. Appl., vol. 24, no. 1, p p. 343 ± 366, Feb. 2021.
[26] L. S. - ( 3 6 5 H V H D U F K D Q G X Q G H I L Q H G ³ 3 H D N O R D G I R U H F D V W L Q J X V L Q J % D \ H V L D Q
regularization, Resilient and adaptive backpropagation lea rning based artificial neur al
Q H W Z R U N V ´ ( O V H Y L H U $ F F H V sed: Jul. 02, 2021.
[27] G. Li, B. - J. Li, X. - G. Yu, and C. - 7 & K H Q J ³ ( F K R 6 W D W H 1 H W Z R U N Z L W K % D \ H V L D Q 5 H J X O D U L ] D W L R Q
for Forecasting Short - 7 H U P 3 R Z H U 3 U R G X F W L R Q R I 6 P D O O + \ G U R S R Z H U 3 O D Q W V ´ Y R O S S ±
12241, 2015.
[28] K. Jazayeri, M. Jazayeri, S. U. - I. C. R Q ' 0 L Q L Q J D Q G X Q G H I L Q H G ³ D Q D O \ V L V R I / H Y H Q E H U J -
Marquardt and Bayesian regularization backpropagation alg orithms in photovoltaic power
H V W L P D W L R Q X V L Q J D U W L I L F L D O Q H X U D O Q H W Z R U N ´ 6 S U L Q J H U $ F F H V V H G - X O
[29] M. H. Alomari, O. Younis, and S. 0 $ + D \ D M Q H K ³ $ 3 U H G L F W L Y H 0 R G H O I R U 6 R O D U 3 K R W R Y R O W D L F
Power using the Levenberg - Marquardt and Bayesian Regulariza tion Algorithms and Real - Time
: H D W K H U ' D W D ´ $ F F H V V H G - X O
[30] = < X H = 6 R Q J ] K H Q J « / 7 - I. C. on, and undefined 2011, ³ % D \ H V L D Q U H J X O D U L ] D W L R Q % 3
Neural Network model for predicting oil - J D V G U L O O L Q J F R V W ´ L H H H [ S O R U H L H H H R U J $ F F H V V H G - X O
02, 2021.
[31] : : D Q J 3 + $ - 0 9 D Q * H O G H U D Q G - . 9 U L M O L Q J ³ & R P S D U L Q J % D \ H V L D Q U H J X O D U L ] D W L R Q D Q G
cross - validated early - stop S L Q J I R U V W U H D P I O R Z I R U H F D V W L Q J Z L W K $ 1 1 P R G H O V ´ , $ + 6 - AISH Publ.,
no. 311, pp. 216 ± 221, 2007.
[32] 5 1 + L G D \ D W 5 5 , V Q D Q W R D Q G 2 ' 1 X U K D \ D W L ³ , P S O H P H Q W D V L - D U L Q J D Q 6 \ D U D I 7 L U X D Q
Perambatan Balik untuk Memprediksi Harga Logam Mulia Emas Menggunakan Alg oritma
/ D Y H Q E H U J 0 D U T X D U G W ´ - 7 H N Q R O G D Q 6 L V W . R P S X W Y R O Q R S
[33] 0 7 0 + 6 $ - $ 0 ) + D Q L I I ³ 0 D Q L I R O G D E V R O X W H S U H V V X U H H V W L P D W L R Q X V L Q J Q H X U D O Q H W Z R U N
Z L W K K \ E U L G W U D L Q L Q J D O J R U L W K P B ( Q K D Q F H G 5 H D G H U S G I ´
[34] I. , Q W D Q 5 L V P D \ D Q L 6 $ ' * K D Q L 1 X U G L Q D Q G $ 7 & . R V Z D U D ³ 3 H U I R U P D Q F H $ Q D O \ V L V R I
: H D W K H U ) R U H F D V W L Q J X V L Q J 0 D F K L Q H / H D U Q L Q J $ O J R U L W K P V ´ - 3 H N R P P D V Y R O Q R S S ±
8, 2021.