Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management

Surakarta, Indonesia, September 14-16, 2021

Support Vector Regression (SVR) Model for Seasonal Time

Series Data
Hanifah Muthiah, Umu Sa’adah, Achmad Efendi
Department of Statistics, Brawijaya University Malang, Indonesia
[email protected], [email protected], [email protected]

Abstract

Support Vector Regression (SVR) is one of the methods used in the supervised learning process for regression cases
that comes from the Support Vector Machine (SVM). SVR is a method to solve forecasting cases that can overcome
overfitting so that it will produce a good performance and has an advantage in optimization with good generalizability
and accuracy results. There are several choices of kernel functions that can be seen from its ability to work on the
SVR method, one of the most popular and often considered the best is the Radial Basis Function (RBF) because based
on some previous research this kernel shows the lowest error value compared to other kernels so that the purpose of
this study is to compare the RBF kernel with other kernels to see the performance of the model and the accuracy of
forecasting produced by using the kernel on time series data which has a seasonal pattern.

Keywords
Kernels, Seasonal Time Series, Support Vector Regression.

1. Introduction
Data Mining is a work process using one or more computer learning techniques (machine learning) to convert large
amounts of data into information or knowledge (Vijayakumar & Nedunchezhian, 2012). The main purpose of data
mining is to make it easier to extract data and make decisions in an analysis (Han et al., 2012). There are several data
mining methods, including predictive ones, namely classification and, regression which aim to predict the value of a
variable based on the value of other variables. The regression technique is a data mining technique that is used to
predict the value of a variable based on the value of another variable, where the predictor variable is a known attribute,
while the response variable is the value to predict (Scholkopf & Smola, 2018).

One of the methods used in the supervised learning process for regression is Support Vector Regression (SVR).
Support Vector Regression is a Support Vector Machine (SVM) method use in regression cases with a large number
of data inputs. SVM is a very popular artificial intelligence-based method. SVM tries to find the best hyperplane
(dividing line) between classes. Various researches have been carried out by applying the SVR method to solve
forecasting cases with good accuracy compared to other methods. SVR is a method that can overcome overfitting so
that it will produce a good performance and have advantages in optimizing the pattern recognition system with good
generalization and accuracy results (Bharati & Ramageri, 2010).

Time series modeling has become popular in recent years. Using historical information to characterize and forecast
the series (Caraka et al., 2017). Prediction or forecasting is an important method in carrying out effective and efficient
planning to obtain future predictions based on the past, controlling the process of producing data series, and
understanding the mechanism of producing the series. Forecasting and planning need to be done because of the grace
period between an event and an upcoming event.

Data is an important thing that supports the test of a predicting or forecasting model. The type of data selected is
determined based on the method to be used. In the prediction or forecasting method, time-series data are generally
used. In time series, there are several types of data patterns, one of which is seasonal patterns. This pattern is a
fluctuation of data that occurs periodically within a certain time, such as in one year, quarter, quarterly, monthly,
weekly, or daily. In general, seasonal patterns happen in recurring conditions that tend to be regular, for example, an
increase in the number of tourists near the end of the year or new year (Percival & Walden, 2000).

© IEOM Society International 3191

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

Support Vector Regression (SVR) as a development method for SVM has a reliable performance in predicting or
forecasting time series data (Hanke & Wichern, 2009). SVR is successfully used in various research fields, such as
prediction on stock price movement data (Henrique et al., 2018) (Patriya, 2020), prediction of sales data (Wu, 2017)
(Nava et al., 2018), and many more. Several studies have shown that the performance of the SVR is better than the
artificial neural network (ANN) model in prediction (Barbour et al., 2018) (Mustakim et al., 2016) and the reliability
of SVR performance is largely determined by the kernel function used and the characteristics of the data used in
constructing the SVR (Caraka et al., 2017) (Alida & Mustikasari, 2020). The best kernel according to several studies
that have been done is the radial basis function (RBF) because it shows the smallest error value compared to other
kernels (Caraka et al., 2017) (Alida & Mustikasari, 2020).

2. Literature Review

2.1 Previous Research

Research (Barbour et al., 2018), which predicts travel time on an urban arterial road in Chennai, India along 2.8 km
by determining 8 optimal input values, a C value of 25, and an error of 0.1. The results obtained show that the SVR
capability is better than the artificial neural network (ANN) model and the moving average with a mean absolute
percentage error (MAPE) of 10 and a root mean squared error (RMSE) of 95%. Research (Alida & Mustikasari, 2020),
regarding the prediction of the rupiah exchange rate towards the dollar, the prediction was done by one of the machine
learning methods, namely the Support Vector Regression (SVR) algorithm. The prediction model uses 3 kernels. Each
kernel is calculated the best model accuracy value and error value for later comparison. In terms of accuracy and error,
the RBF kernel has advantages over linear and polynomial kernels with the coefficient of determination (R-squared)
of 95.94% and the root mean squared error (RMSE) of 1.25%. Research (Bagheri & Rezaei, 2019), predicts
permeability (the ability of fluids to pass through porous media without changing the rock structure that plays a major
role in the production rate of a hydrocarbon reservoir) in reservoir rock.

The method used is Support Vector Regression (SVR) with a kernel radial basis function (RBF) to estimate
permeability in Iran's South Pars gas field. The scores for the four electrofacies appointed were 88.2, 78.51, 84.73,
and 77.54 percent, respectively. The high value of the coefficient of determination (R-Squared) obtained for each
electrofacies shows reliable accuracy in the model for predicting reservoir permeability. Research (Chawsheen &
Broom, 2017), analyzes the performance of various models constructed using linear kernel SVR and trained on
historical bid data for high-frequency currency trading. This model generates a simulation of currency trading in the
following year. This model is used to execute record profit, hit ratio, and the number of trades, with the result that
it is possible to get good profit as well as hit ratio from the trained linear model. Based on descriptions of some of
the previous research literature that became the reference for this study, the application of the Support Vector
Regression (SVR) algorithm with the radial basis function (RBF) and Linear kernel was carried out in prediction and
forecasting on the seasonal time series Support Vector Regression (SVR) data. Data with seasonal patterns have
special characteristics that can be applied to several predictions or forecasting methods to find out how accurate the
model is.

2.2. Theoretical Basis

2.2.1 Time Series

Time series is a set of observations of ordered data in time. The time series method is a method of forecasting
using an analysis of the pattern of the relationship between the variables to be estimated and the time variable.
Forecasting a time series data needs to pay attention to the type or pattern of data. In general, there are four kinds of
time series data patterns, namely horizontal, trend, seasonal and cyclical (Hanke & Wichern, 2009). The four patterns
in time series are the trends, cyclic, seasonal, and random variation (irregular). Seasonal data patterns occur when
time-series data is influenced by seasonal factors. This pattern has a seasonal pattern that repeats itself from period to
period. For example, there are repeated fluctuations at certain times, such as what is often found in the quarterly,
semester, monthly, or weekly data.

© IEOM Society International 3192

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

2.2.2 Support Vector Regression

Support Vector Regression (SVR) is an SVM development for regression cases. In the case of regression, the output
is either a real number or a continuous number. SVR is a method that can overcome overfitting so that it will produce
good performance (Scholkopf & Smola, 2018). If the SVM aims to divide the dataset (classification) into two zones
(clusters), then SVR is the opposite, which is to make the entire dataset into one zone, while maximizing the epsilon
distance (𝜀𝜀) (small value). You can see the SVR illustration in Figure 1.

Figure 1. SVR Illustration (Babu & Mohan, 2017)

Figure 1 shows a hyperplane (diagonal line in the middle) that is flanked by two lines + and a boundary -. It also can
be seen that there is 𝜀𝜀 as the distance between the hyperplane and the two boundary lines. There are several data points
circled which become potential support vectors, meaning that these data points are data points that can become
potential boundaries. The purpose of the SVR is to find a function as a hyperplane (dividing line) in the form of a
regression function that matches all data input with the smallest possible error (Scholkopf & Smola, 2018). The
purpose of this SVR is to map the input vector into a higher dimension. Suppose a function (4) is a regression line as
the optimal hyperplane (Abe, 2005).
𝑓𝑓(𝑥𝑥) = 𝒘𝒘𝑇𝑇 𝜑𝜑(𝑥𝑥) + 𝑏𝑏 (1)
𝑤𝑤 is the weight vector 𝑙𝑙 dimension, 𝜑𝜑(𝑥𝑥) is a function that maps 𝑥𝑥 to the feature space with 𝑙𝑙 dimensions, 𝑏𝑏 is biased.
The function of 𝜑𝜑(𝑥𝑥) shows a point in the higher dimensional feature space, the result of the mapping of the input
vector in the lower dimensional input space. The coefficients 𝑤𝑤 and 𝑏𝑏 are estimated by minimizing the risk function.
Therefore, to maximize the margin 𝛿𝛿, a minimum ‖𝑤𝑤‖ is needed. Optimization of problem-solving as shown in
function (2).
𝑚𝑚𝑚𝑚𝑚𝑚12‖𝑤𝑤‖² (2)
with the condition, 𝑦𝑦𝑖𝑖 − 𝑤𝑤 𝑇𝑇 𝜑𝜑(𝑥𝑥𝑖𝑖 ) − 𝑏𝑏 ≤ 𝜀𝜀, 𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖 = 1, … , 𝑙𝑙 and 𝑤𝑤 𝑇𝑇 𝜑𝜑(𝑥𝑥𝑖𝑖 ) − 𝑦𝑦𝑖𝑖 + 𝑏𝑏 ≤ 𝜀𝜀, 𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖 = 1, … , 𝑙𝑙.
Where 𝑦𝑦𝑖𝑖 is the actual value of 𝑖𝑖 period, and 𝜑𝜑(𝑥𝑥𝑖𝑖 ) is the estimated value of 𝑖𝑖 period.

The ‖𝑤𝑤‖² factor is called regulation. Minimizing ‖𝑤𝑤‖² will make a function as thin (flat) as possible so that it can
control the functional capacity. All points outside the margin/limit 𝜀𝜀 will be penalized. Furthermore, the optimization
problem above can be formulated into a function (3) (Scholkopf & Smola, 2018).
𝑚𝑚𝑚𝑚𝑚𝑚12‖𝑤𝑤‖² + ∁ ∑ℓ𝑖𝑖=1(𝜉𝜉𝑖𝑖 + 𝜉𝜉𝑖𝑖∗ ) (3)
with the condition, 𝑦𝑦𝑖𝑖 − 𝑤𝑤 𝑇𝑇 𝜑𝜑(𝑥𝑥𝑖𝑖 ) − 𝑏𝑏 − 𝜉𝜉𝑖𝑖 ≤ 𝜀𝜀, 𝑖𝑖 = 1, … , 𝑙𝑙
𝑤𝑤 𝑇𝑇 𝜑𝜑(𝑥𝑥𝑖𝑖 ) − 𝑦𝑦𝑖𝑖 + 𝑏𝑏 − 𝜉𝜉𝑖𝑖∗ ≤ 𝜀𝜀, 𝑖𝑖 = 1, … , 𝑙𝑙
𝜉𝜉𝑖𝑖 , 𝜉𝜉𝑖𝑖∗ ≥ 0
𝐶𝐶 > 0 constant determines how much the error deviation is from the tolerable limit 𝜀𝜀. The formula above is a Convex
Linear Programming NLP Optimization Problem which functions to minimize the quadratic function to be converted
into a constraint. This limitation can be solved by using the Lagrange Multiplier function. The process of deriving
formulas is very long and complicated. After going through mathematical stages, a new equation is obtained with the
function:
𝑓𝑓(𝑥𝑥) = ∑𝑙𝑙𝑖𝑖=1(𝑎𝑎𝑖𝑖 − 𝑎𝑎𝑖𝑖∗ ). (𝑥𝑥𝑖𝑖 . 𝑥𝑥) + 𝑏𝑏 (4)

© IEOM Society International 3193

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

Where 𝑥𝑥𝑖𝑖 is the support vector and 𝑥𝑥 is the test vector. The above functions can be used to solve linear problems.
Whereas for non-linear problems the values of 𝑥𝑥𝑖𝑖 , and 𝑥𝑥 are first transformed into a high-dimensional feature space
by mapping the vectors 𝑥𝑥𝑖𝑖 and 𝑥𝑥 into the kernel function so that the final function becomes:
𝑓𝑓(𝑥𝑥) = ∑𝑙𝑙𝑖𝑖=1(𝑎𝑎𝑖𝑖 − 𝑎𝑎𝑖𝑖∗ ). 𝐾𝐾(𝑥𝑥𝑖𝑖 . 𝑥𝑥) + 𝑏𝑏 (5)
The function 𝐾𝐾(𝑥𝑥𝑖𝑖 . 𝑥𝑥) is the Kernel. The table 1 below shows the kernels used in the SVR calculation (Scholkopf &
Smola, 2018).
Table 1. Kernel Function

Kernel Function
Linear 𝐾𝐾(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 ) = (𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 )
Radial Basis Function ∥𝑥𝑥⃗𝑖𝑖 −𝑥𝑥⃗𝑗𝑗 ∥2
𝐾𝐾(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 ) = 𝑒𝑒𝑒𝑒𝑒𝑒(−
2𝜎𝜎²
)

2.2.3 Goodness of Fit Models

The measure of model quality is used to evaluate the model that has been formed as well as to see the effectiveness of
a forecast. The measure used in this study is the root mean squared error (RMSE).
∑𝑛𝑛 ̇ 2
𝑖𝑖=1(𝑌𝑌𝑖𝑖 −𝑌𝑌𝑖𝑖 )
𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹 = √𝑴𝑴𝑴𝑴𝑴𝑴 = � (6)
𝑛𝑛
Where 𝑌𝑌𝑖𝑖 is the actual data, 𝑌𝑌𝑖𝑖̇ is the forecast result data, and 𝑛𝑛 is the number of observed data (Voulgaraki, 2013).

3. Methods
This paper uses two processes, namely pre-processing and post-processing. pre-processing is meant here is to form
input 𝒙𝒙 using lags from SARIMA to then carry out Support Vector Regression (SVR) analysis as a post-process, or
it can be called the main analysis. Based on Figure 2, this is the explanation detail of the research:
Check data training pattern.
Stationary test with ADF, if data is not stationary, do a difference for data training, and then check again with ADF
test.
Check ACF and PACF patterns.
SARIMA parameter estimation for training data.
Testing the suitability of the SARIMA model on the residuals, namely the white noise and normality.
If testing the suitability residual data is not significant, forming the SARIMA model with input data using stepwise
and no-stepwise mode for Linear and Radial Basis Function (RBF) kernels.
Calculate Root Mean Squared Error (RMSE) for each SVR model.
Validation of each model with testing data.
Calculate Root Mean Squared Error (RMSE) for testing data.
Compare the RMSE of each SVR model to choose the best model.

© IEOM Society International 3194

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

Start

Training Data Pattern

Stationary Test by ADF

Is The Data
Yes No
Stationary?

Differencing
ACF Pattern The Data
and
PACF Pattern

SARIMA Parameter
Is Residual Normal?
Estimation for Training Data
No

Is Residual White Forming The SARIMA Model:

Yes
Yes Noise? 1. Input data with the SARIMA model
based on stepwise using Linear and
RBF kernel function.
2. Input data with the SARIMA model
Calculate RMSE of SARIMA No based on no-stepwise using Linear and
Model for Training Data RBF kernel function.

Calculate RMSE for

Validation of SARIMA Each SVR Models
Model with Testing Data

Validation of Each Models

Calculate RMSE of Testing Data with Testing Data

Compare RMSE of each SVR Models

Finish Calculate RMSE of Testing Data
to Choose The Best Model

Figure 2. Research Flow Chart

4. Data Collection
In this case, secondary data were used. Secondary data uses a sample of data on the realization of daily electricity
loads in the East Java region in units of megawatts (MW) in the period October 07, 2019, to April 05, 2020, obtained
from the Java-Bali State Perusahaan Listrik Negara (PLN). The data length 𝑛𝑛 = 182 data divided into training data
as much as 161 data from the number of samples used which aims to form a model and testing data as many as 21
data from the number of samples used for model testing to predict and forecast data. The data type is weekly seasonal
it can be seen in Figure 3.

5. Result and Discussion

5.1 Numerical Results

5.1.1 Support Vector Regression (SVR) with Stepwise Input

In Table 2, there is RMSE as a measure of the goodness of the SVR model with stepwise input for two kernels were
chosen, namely Linear function and Radial Basis Function (RBF).

© IEOM Society International 3195

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

Table 2. Summary SVR Models with Stepwise Input

Kernel RMSE
Linear Function 204.0512
Training
Radial Basis Function 209.3073
Linear Function 153.8064
Testing
Radial Basis Function 159.9443

Based on Table 2, the linear function kernel has the smallest RMSE value for training data is 204.0512, and testing
data is 153.8064, while RBF kernel value for training data is 209.3073 and testing data is 159.9443. The value obtained
does not differ much between these two kernels.

5.1.2 Support Vector Regression (SVR) with No-Stepwise Input

In Table 3, there is RMSE as a measure of the goodness of the SVR model with no-stepwise input for two kernels
were chosen, namely Linear function and Radial Basis Function (RBF).

Table 3. Summary SVR Models with No-Stepwise Input

Kernel RMSE
Linear Function 284.2707
Training
Radial Basis Function 286.9821
Linear Function 234.0197
Testing
Radial Basis Function 232.1223

Based on Table 3, the linear function kernel has the smallest RMSE value for training data is 284.2707, and RBF for
testing data is 232.1223, while RBF kernel value for training data is 286.9821 and linear function for testing data is
234.0197. Same with SVR with stepwise input, the value obtained does not differ much between these two kernels
with no-stepwise input.

5.2 Graphical Results

5.2.1 Secondary Data

At this stage, the data pattern is checked using a line chart. The expected data criterion is a seasonal pattern. An
overview of the data patterns used in this study can be seen in Figure 3. Electricity load realization data for October
07, 2019, to April 05, 2020, tends to be stationary with a seasonal trend at the end of the week, namely from Saturday
to Sunday, and then it will rise again on Monday. Generally, this happens because Monday to Friday are working days
where more activity is carried out so that the load on electricity consumption increases more than on weekends.

© IEOM Society International 3196

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

REALIZATION OF ELECTRICITY LOAD IN EAST JAVA

(MEGA WATT)
7000
6000
5000
4000
3000
2000
1000
0

TUE, 4 FEB

SUN, 15 MAR

SAT, 4 APR
FRI, 1 NOV

SAT, 16 NOV

MON, 16 DES

MON, 24 FEB

FRI, 20 MAR
WED, 6 NOV

THU, 21 NOV
TUE, 26 NOV

TUE, 10 MAR

WED, 25 MAR
MON, 7 OKT

FRI, 14 FEB
SUN, 27 OKT

MON, 11 NOV

SUN, 1 DES
FRI, 6 DES

SUN, 5 JAN
FRI, 10 JAN
WED, 15 JAN

SAT, 25 JAN

SAT, 29 FEB
THU, 5 MAR

MON, 30 MAR
WED, 11 DES

SAT, 21 DES

THU, 30 JAN

SUN, 9 FEB

WED, 19 FEB
SAT, 12 OKT

THU, 26 DES
TUE, 31 DES

MON, 20 JAN
THU, 17 OKT
TUE, 22 OKT

Figure 3. Realization of Electricity Load in East Java

5.2.2 Seasonal Autoregressive Integrated Moving Average (SARIMA)

Significant SARIMA lags are used for SVR input, and seasonal lag enters the input. In this case, using stepwise and
no-stepwise mode. In Figure 4 is the ACF pattern for SARIMA, lag shows that are a seasonal pattern, which is similar
to Figure 3.

Figure 4. ACF of SARIMA

Based on ACF pattern, SARIMA using stepwise is SARIMA (1,0,0)(2,0,0)[7], and for no-stepwise is SARIMA
(3,0,0)(2,0,0)[7].

5.2.3 Result of SVR Linear Kernel with Stepwise Input

The results of SVR forecasting on testing data using the linear kernel function with stepwise input can be seen in
Figure 5. It can be seen that the forecasting results and testing / actual data have a pattern that tends to be the same.
Testing data is used to test the resulting model from the training data. Training and testing can see the similarity or
suitability of their characteristics based on the closeness of the results of the goodness of the models. testing data used
starts from March 16, 2020, to April 05, 2020. Forecasting is carried out for 3 weeks or 3 seasonal periods, from April
06 to April 26, 2020.

© IEOM Society International 3197

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

TESTING VS FORECAST
Testing Forecast

5800
5600
ELECTRICITY LOAD

5400
5200
(MW)

5000
4800
4600
4400
4200

Figure 5. Testing vs Forecast

5.4 Validation

5.4.1 SVR Model with Stepwise Input

Linear function and RBF kernels generate parameter values of 𝐶𝐶 = 1, 𝛾𝛾 = 0.3333333, 𝑎𝑎𝑎𝑎𝑎𝑎 𝜀𝜀 = 0.1 with 𝜎𝜎 =
335.0145, produce 𝛽𝛽𝐿𝐿𝐿𝐿𝐿𝐿 = −0.004917671, and 𝛽𝛽𝑅𝑅𝑅𝑅𝑅𝑅 = 0.1108223.

1. Linear Kernel Function

𝐾𝐾(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 ) = (𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 )
𝑓𝑓(𝑥𝑥) = −0.004917671(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 )
2. Radial Basis Function
∥ 𝑥𝑥⃗𝑖𝑖 − 𝑥𝑥⃗𝑗𝑗 ∥2
𝐾𝐾(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 ) = exp(− )
2𝜎𝜎²
2
∥ 𝑥𝑥⃗𝑖𝑖 − 𝑥𝑥⃗𝑗𝑗 ∥
𝐾𝐾(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 ) = exp(− )
2(335.01452 )
∥𝑥𝑥⃗𝑖𝑖 −𝑥𝑥⃗𝑗𝑗 ∥2
𝑓𝑓(𝑥𝑥) = 0.1108223(exp(− ))
2(335.01452 )

5.4.2 SVR Model with No-Stepwise Input

Linear function and RBF kernels generate parameter values of 𝐶𝐶 = 1, 𝛾𝛾 = 0.2, 𝑎𝑎𝑎𝑎𝑎𝑎 𝜀𝜀 = 0.1 with 𝜎𝜎 = 343.0039,
produce 𝛽𝛽𝐿𝐿𝐿𝐿𝐿𝐿 = −0.01964134, and 𝛽𝛽𝑅𝑅𝑅𝑅𝑅𝑅 = 0.03323177.

1. Linear Kernel Function

𝐾𝐾(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 ) = (𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 )
𝑓𝑓(𝑥𝑥) = −0.01964134(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 )
2. Radial Basis Function
∥ 𝑥𝑥⃗𝑖𝑖 − 𝑥𝑥⃗𝑗𝑗 ∥2
𝐾𝐾(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 ) = 𝑒𝑒𝑒𝑒𝑒𝑒(− )
2𝜎𝜎²
2
∥ 𝑥𝑥⃗𝑖𝑖 − 𝑥𝑥⃗𝑗𝑗 ∥
𝐾𝐾(𝑥𝑥⃗𝑖𝑖 , 𝑥𝑥⃗𝑗𝑗 ) = 𝑒𝑒𝑒𝑒𝑒𝑒(− )
2(343.00392 )

© IEOM Society International 3198

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

∥ 𝑥𝑥⃗𝑖𝑖 − 𝑥𝑥⃗𝑗𝑗 ∥2
𝑓𝑓(𝑥𝑥) = 0.03323177(𝑒𝑒𝑒𝑒𝑒𝑒(− ))
2(343.00392 )

5.4.3 Goodness of Fit SVR Model

Based on the previous Table 2 and Table 3, it is found that the linear kernel function with stepwise input has the best
model accuracy for RMSE, which means 153.8064 for testing and 204.0512 for training invariance of the Realization
of Electricity Load in East Java.

6. Conclusion
Based on the analysis that has been done, the conclusions are obtained, namely:
1. Based on the RMSE value obtained from the linear kernel function and RBF, it was found that the linear kernel
function using stepwise for data testing had the smallest RMSE value, namely 153.8064.
2. Based on the RMSE value obtained, linear kernel using stepwise has the best performance and accuracy in
prediction, so that it can be said that the SVR model is simple for Realization of Electricity Load in East Java
data.

References
Abe, S. (2005). Support vector machines for pattern classification (Vol. 2). Springer.
Alida, M., & Mustikasari, M. (2020). Rupiah Exchange Prediction of US Dollar Using Linear, Polynomial, and
Radial Basis Function Kernel in Support Vector Regression. Jurnal Online Informatika, 5(1), 53–60.
Babu, N. R., & Mohan, B. J. (2017). Fault classification in power systems using EMD and SVM. Ain Shams
Engineering Journal, 8(2), 103–111.
Bagheri, M., & Rezaei, H. (2019). Reservoir rock permeability prediction using SVR based on radial basis function
kernel. Carbonates and Evaporites, 34(3), 699–707.
Barbour, W., Mori, J. C. M., Kuppa, S., & Work, D. B. (2018). Prediction of arrival times of freight traffic on US
railroads using support vector regression. Transportation Research Part C: Emerging Technologies, 93, 211–
227.
Bharati, M., & Ramageri, M. (2010). Data mining techniques and applications.
Caraka, R. E., Yasin, H., & Basyiruddin, A. W. (2017). Peramalan Crude Palm Oil (CPO) Menggunakan Support
Vector Regression Kernel Radial Basis. Jurnal Matematika, 7(1), 43–57.
Chawsheen, T. A., & Broom, M. (2017). Seasonal time-series modeling and forecasting of monthly mean
temperature for decision making in the Kurdistan Region of Iraq. Journal of Statistical Theory and Practice,
11(4), 604–633.
Han, J., Kamber, M., & Pei, J. (2012). Data mining: concepts and techniques, Waltham, MA. Morgan Kaufman
Publishers, 10, 971–978.
Hanke, J. E., & Wichern, D. W. (2009). Business forecasting 9th ed. New Jersey.
Henrique, B. M., Sobreiro, V. A., & Kimura, H. (2018). Stock price prediction using support vector regression on
daily and up to the minute prices. The Journal of Finance and Data Science, 4(3), 183–201.
Mustakim, M., Buono, A., & Hermadi, I. (2016). Performance comparison between support vector regression and
artificial neural network for prediction of oil palm production. Jurnal Ilmu Komputer Dan Informasi, 9(1), 1–8.
Nava, N., Di Matteo, T., & Aste, T. (2018). Financial time series forecasting using empirical mode decomposition
and support vector regression. Risks, 6(1), 7.
Patriya, E. (2020). Implementasi Support Vector Machine Pada Prediksi Harga Saham Gabungan (IHSG). Jurnal
Ilmiah Teknologi Dan Rekayasa, 25(1), 24–38.
Percival, D. B., & Walden, A. T. (2000). Wavelet methods for time series analysis (Vol. 4). Cambridge university
press.
Scholkopf, B., & Smola, A. J. (2018). Learning with kernels: support vector machines, regularization, optimization,
and beyond. Adaptive Computation and Machine Learning series.
Vijayakumar, V., & Nedunchezhian, R. (2012). A study on video data mining. International Journal of Multimedia
Information Retrieval, 1(3), 153–172.
Voulgaraki, M. (2013). Forecasting sales and intervention analysis of durable products in the Greek market.
Empirical Evidence from the New Car Retail Sector. London School of Economics and Political Science.

© IEOM Society International 3199

Proceedings of the Second Asia Pacific International Conference on Industrial Engineering and Operations Management
Surakarta, Indonesia, September 14-16, 2021

Wu, J. Y. (2017). Housing Price prediction Using Support Vector Regression.

Biographies

Hanifah Muthiah is a Student of the Magister Program in Statistics Department at Brawijaya University Malang.
She received a bachelor's degree in statistics from the Islamic University of Indonesia Yogyakarta. She is interested
in the scientific field of statistics, especially in Regression Analysis.

Umu Sa’adah is Lecturer of Mathematics Department at Brawijaya University Malang. She has expertise in Statistical
Computation.

Achmad Efendi is A Lecturer and Head of the Undergraduate Program of Statistics Department at Brawijaya
University Malang. He has expertise in Statistical Modeling.

© IEOM Society International 3200