2022 3rd International Conference on Clean and Green Energy Engineering

Comparison of Medium-Term Load Forecasting

Methods (Splitted Linear Regression and Artificial
Neural Networks) in Electricity Systems Located in
Tropical Regions
Agus Setiawan Zainal Arifin Budi Sudiarto Fauzan Hanif Jufri
2022 3rd International Conference on Clean and Green Energy Engineering (CGEE) | 978-1-6654-5265-6/22/$31.00 ©2022 IEEE | DOI: 10.1109/CGEE55282.2022.9976521

Electrical Engineering Engineering and Technology Electrical Engineering Electrical Engineering

Department Division Department Department
Universitas Indonesia PLN Universitas Indonesia Universitas Indonesia
Jakarta, Indonesia Jakarta, Indonesia Jakarta, Indonesia Jakarta, Indonesia
[email protected] [email protected] [email protected] [email protected]

Qasthalani Haramaini Iwa Garniwa

Electrical Engineering Electrical Engineering
Department Department
Universitas Indonesia Universitas Indonesia
Jakarta, Indonesia Jakarta, Indonesia
[email protected] [email protected]

Abstract—Load forecasting for the medium term to supply forecasting, and long-term load forecasting. Short-term
power plants within 1 month, can optimize the economic dispatch forecasting aims to optimize the supply of electricity in a system
of generators that will be used to supply large systems. The by paying attention to aspects of system reliability and also
accuracy of forecasting the half-hourly load will result in a more economic aspects. The duration generated in this forecast is
efficient electricity supply in a system that uses flat rates on the generally in half an hour or 1 hour, this forecast lasts for 1 day
utility side. In this paper, the author tries to compare the to weekly. Meanwhile, medium-term load forecasting aims to
forecasting methods of linear regression, Artificial Neural plan fuel reserves and other primary energy and to confirm unit
Networks, and Splitted Linear Regression. This method is applied commitments. The results of this forecast are generally in the
to the largest system in Indonesia, a country located at the tropical
form of peak load and average daily consumption. Forecasting
region which has different characteristics from countries that have
four seasons. At the end of this study, it can be concluded that the
duration from 1 month to 1 year ahead. And lastly is the long-
Splitted Linear Regression method has the highest performance term load forecasting, this forecast is used for expansion of the
with the lowest MAPE value of 2.63%. construction of new power plants and also transmission system
used for power evacuation. The duration for this forecast lasts
Keywords—medium-term load forecasting, Splitted linear for the estimated load from annual to 10 years.
regression, MAPE, deep learning load forecasting Researchers over the past few decades have used many
I. INTRODUCTION methods to perform load forecasting. Load forecasting methods
are generally divided into two, namely statistical methods and
Electricity is a human need in everyday life and one of the artificial intelligent methods. Statistical forecasting methods are
main driving factors for a country's economy growth. In used if there is data that can be taken from research in the form
providing adequate electricity to meet electricity needs at any of time series data[2]. Time series data is past data consisting of
time reliably, it is necessary to estimate the load for planning the results of measurements or observations arranged in series
electricity production. Because the electricity supply has an according to the order of time. The more data that can be
uneasy process to manage the selection of appropriate and obtained for several types of statistical methods, the more
inexpensive generators, maintenance schedule arrangements accurate the forecasting results will be.
and moreover about plant planning due to electricity demand
growth. Medium term load forecasting (MTLF) is needed to Statistical methods consist of several methods including
address the needs for optimizing the supply of electricity. time series techniques, linear regression [3], linear
autoregressive methods [4], exponential methods [5] and also
A simple classification of electrical load forecasting has includes the stochastic time series method[6]. Also included are
been presented in this paper [1] and from this paper load Autoregressive Integrated Moving Average (ARIMA) [7, 8],
forecasts can be divided into 3 types based on the forecast Autoregressive (AR) [9]. The effect of sudden changes in
period, namely short-term load forecasting, medium-term load environmental or sociological variables such as changes in

978-1-6654-5265-6/22/$31.00 ©2022 IEEE 84

Authorized licensed use limited to: Indian Institute of Technology Gandhinagar. Downloaded on August 28,2024 at 14:09:16 UTC from IEEE Xplore. Restrictions apply.
weather, holidays, the errors that arise will become larger and
this can be observed [10]. This is the main point obtained from
the statistical method [11]. Artificial intelligence methods
include artificial neural networks (ANN) [12-15], recurrent
neural networks, deep learning [16, 17], LSTM [17, 18], fuzzy
systems [19], knowledge based expert systems (KBES), wavelet
analysis, support vector machines (SVM), nd so on.
This paper presents medium-term load forecasting that
compares the time series split regression method, linear
regression time series and also deep learning methods in matlab
program for time series. The data used is a time series of
approximately 900 days on the Java-Bali system. The Java-Bali
system is the largest system in Indonesia which is the center of
the country's economic growth. This system is located at the
equator, so the daily load curve is relatively the same throughout
the year. In contrast to systems located in sub-tropical areas Fig. 1. Data load profile sistem Jawa Madura Bali
which have 4 seasons, so they have high electricity consumption
differences depending on the season. As in paper [16] shows B. Method
different load patterns throughout the year in Korea or in paper By paying attention to the characteristics of the daily load
[20] which shows different patterns throughout the year in profile on the system in countries located in the equatorial region
Australia. which tend to be the same throughout the year. As shown in
The model will estimate the load for 1 month so it is Figure 1 above. Researchers see differences in daily load
classified as a medium-term load forecasting type. However, the profiles occur due to differences in weekdays and weekends as
advantage of this paper is that although the method used is well as holiday events in the system. Then the author takes this
included in the medium-term load forecasting category which as an initial hypothesis and proposes a load forecasting method
has the characteristics of the results in the form of peak loads or as shown in Figure 2 as follows:
daily energy consumption only, the results of each model are in
the form of half-hour data that can be used as short-term load
forecasting. . That is one of the uniqueness of this article. At the
end, the comparison of forecasting accuracy for the three
methods above is presented in the Mean Average Percentage
Error (MAPE) [21] and also the Root Mean Square Error
(RMSE) [22].
This paper is structured as follows. In part II the method used
to obtain medium-term load forecasting. Section III presents the
results obtained from each of the methods compared in this
paper. In part IV about the discussion of the results. Finally,
conclusions are drawn in section V.
II. MATERIAL & METHOD
A. Load Profile Data
The profile of the use of electrical energy for each system
has different characteristics. If we compare the electrical load
profile for countries located in the sub-tropics, they tend to have
different variations in use throughout the year following the
changing seasons. For example, use in winter tends to consume
more electricity to fulfill room heating. Countries located in the
tropics generally have a relatively similar electrical load profile
throughout the year. Indonesia is a country located in the tropics
of the earth so in general, electric power systems in Indonesia
have the same characteristics throughout the year. In this study,
the authors present the data obtained for approximately 900 days
of load profile data on the largest system in Indonesia. The data Fig. 2. Research methodology
can be seen in Figure 1.
By looking at the methodology in Figure 2, the author would
like to explain the 3 methods that will be used in this research as
follows:

85
Authorized licensed use limited to: Indian Institute of Technology Gandhinagar. Downloaded on August 28,2024 at 14:09:16 UTC from IEEE Xplore. Restrictions apply.
 1) Splitted linear regression method
This method is a simple linear method using time series data
obtained from electrical loads that are recorded for
approximately 900 days. The simple linear equation method is
as follows:
𝑌 = 𝑎 + 𝑏𝑋 (1)

(∑ 𝑦)(∑ 𝑥 2 ) − (∑ 𝑥)(∑ 𝑥𝑦)

𝑎= (2)
𝑛(∑ 𝑥 2 ) − (∑ 𝑥)2
𝑛(∑ 𝑥𝑦) − (∑ 𝑥)(∑ 𝑦)
𝑏= (3)
𝑛(∑ 𝑥 2 ) − (∑ 𝑥)2 Fig. 3. Observation data and forecasting results at 00:30 using Splitted Linear
Regression Method
where:
Y = Forecasting result at time x
X = time x of the time series to be predicted
a = regression constant
b = regression coefficient
n = amount of existing data
x = time x time-series data
y = system load by time x time series data
The existing data has a time interval of half an hour for 24
hours a day. There are 48 time series for each time step. This
Fig. 4. Observation data and forecasting results at 00:30 using the Linear
method pays attention to weekdays and weekends as well as
Regression Method
holidays that exist in the system.
2) Linear regression method
This method is the same as the previous method but does not
consider weekdays and weekends and holidays in the system.
3) Deep learning method
This method belongs to the type of artificially intelligent
method. The author adopts the algorithm provided by Matlab
which is used to predict time series data. This method is a way
of estimating time-series data using a Long Short Term Memory
(LSTM) network. The specified LSTM layer has 200 hidden
units and is trained 250 times.
From the three methods, forecasts will be generated for each
for 30 days. Then from these results, the MAPE and RMSE were Fig. 5. Observation data and forecasting results at 00:30 using Deep Learning
Method
measured and compared with the actual data.
III. RESULT Figures 3, 4, and 5 show the forecasting results of the system
load at 00:30 which consist of 911 observation data and 30
By using the methodology in the previous section, this forecasting data. Observational data is shown in blue, while
section will show the results of load forecasting for each method. forecasting data is shown in orange. Observation and forecasting
Here we can see one of the forecasts for one time step, which is are carried out for 24 hours so that with a half-hour interval, 48
taken at 00:30 for the three models used. data are generated as shown in Figures 3,4, and 5. Since
forecasting is carried out for 30 days, 30 daily load profile
forecasting data is generated. The following is the result of the
daily load curve for the first day of each method.

86
Authorized licensed use limited to: Indian Institute of Technology Gandhinagar. Downloaded on August 28,2024 at 14:09:16 UTC from IEEE Xplore. Restrictions apply.
 IV. DISCUSSION
By looking at the results section, it can be seen that from the
three methods used above, it is found that the Splitted Linear
Regression method has the best MAPE and RMSE values when
compared to the other two methods. If we look at the graph of
the forecasting time series results at the same time within 24
hours, it is as shown in Figures 3,4, and 5. Based on this figure,
it can be seen that both splitted and linear regression methods
tend to look very different from the trend of observational data.
Meanwhile, the deep learning method shows that forecasting
data is more dynamic and looks more like the trend with
Fig. 6. The first day of the daily load curve forecasting results using the observation data. However, if you look at the daily load curve,
Splitted Linear Regression Method for example on the results of the first day of forecasting which
can be seen in Figures 6, 7, and 8, it can be seen that the daily
load curves for the two regression methods are more similar in
shape to the previous observation daily curve data, which can be
seen in Figure 1 Meanwhile, the deep learning method shows a
significant difference compared to the daily load curve of
observations. So when viewed as a whole using the MAPE and
RMSE error measurement methods, the Splitted Linear
Regression Method obtained the best results, which were 2.63%
and 822.84 MW, respectively.
V. CONCLUSION
The Java-Madura-Bali system is an electricity system
located at the equator. So that the results of the method in this
Fig. 7. The first day of the daily load curve forecasting results using the Linear study can be applied to existing systems in the equatorial region.
Regression Method Where systems located at the equator tend to have relatively the
same daily load curve throughout the year. Electricity load in the
equatorial area tends to be influenced by differences in
weekdays and weekends and holidays that exist in the system,
this is indicated by the linear regression method, the separation
between weekdays and weekends and holidays shows better
results than linear regression methods.
By using observational data of about 911 days, it can be
concluded that the best load forecasting method of the 3 methods
mentioned above is the Splitted Linear Regression Method.
With MAPE and RMSE values of 2.63% and 882.84 MW,
respectively.
Further research can be developed by adding methods to be
Fig. 8. The first day of the daily load curve forecasting results using the Deep
compared. So it can be known which load forecasting is the most
Learning Method optimal. This is because load forecasting research in the
equatorial region is still not much compared to research
The last thing to do in this research is to measure the error of conducted in subtropical.
the forecasting results for the three methods above using MAPE
and RMSE. The following are the results of the measurement of ACKNOWLEDGMENT
forecasting results from the three methods. This research was supported by PT PLN (Persero). We thank
all parties involved who provided data, insight, and expertise
TABLE I. MAPE AND RMSE VALUES FOR THE THREE METHODS that greatly assisted the research, as well as comments and
Method MAPE RMSE reviews which helped bring this research to complete fruition.
No
(%) (MW)
1 Splitted Linear Regression Method 2.63 822.84 REFERENCES
[1] K. Metaxiotis, A. Kagiannas, D. Askounis, and J. Psarras, "Artificial
2 Linear Regression Method 5.03 1361.5 intelligence in short term electric load forecasting: a state-of-the-art
survey for the researcher," Energy conversion and Management, vol. 44,
3 Deep Learning Method 7.56 2166.6
no. 9, pp. 1525-1534, 2003.
[2] F. Marín and F. Sandoval, "Short-term peak load forecasting: Statistical
methods versus artificial neural networks," in International Work-
Conference on Artificial Neural Networks, 1997: Springer, pp. 1334-
1343.

87
Authorized licensed use limited to: Indian Institute of Technology Gandhinagar. Downloaded on August 28,2024 at 14:09:16 UTC from IEEE Xplore. Restrictions apply.
[3] N. Amral, C. Ozveren, and D. King, "Short term load forecasting using [14] [3] A. G. Abdullah, W. W. Sopian, W. Arasid, A. B. D. Nandiyanto, A.
multiple linear regression," in 2007 42nd International universities power A. Danuwijaya, and C. U. Abdullah, "Short-term peak load forecasting
engineering conference, 2007: IEEE, pp. 1192-1198. using PSO-ANN methods: The case of Indonesia," Journal of
[4] J.-F. Chen, W.-M. Wang, and C.-M. Huang, "Analysis of an adaptive Engineering, Science, and Technology, vol. 13, no. 8, pp. 2395-2404,
time-series autoregressive moving-average (ARMA) model for short- 2018.
term load forecasting," Electric Power Systems Research, vol. 34, no. 3, [15] [4] H. H. Aly, "A proposed intelligent short-term load forecasting
pp. 187-196, 1995. hybrid models of ANN, WNN and KF based on clustering techniques for
[5] W. Christiaanse, "Short-term load forecasting using general exponential smart grid," Electric Power Systems Research, vol. 182, p. 106191, 2020.
smoothing," IEEE Transactions on Power Apparatus and Systems, no. 2, [16] B.-S. Kwon, R.-J. Park, and K.-B. Song, "Weekly peak load forecasting
pp. 900-911, 1971. for 104 weeks using deep learning algorithm," in 2019 IEEE PES Asia-
[6] Y. Chakhchoukh, P. Panciatici, and L. Mili, "Electric load forecasting Pacific Power and Energy Engineering Conference (APPEEC), 2019:
based on statistical robust methods," IEEE Transactions on Power IEEE, pp. 1-4.
Systems, vol. 26, no. 3, pp. 982-991, 2010. [17] T. Bashir, C. Haoyong, M. F. Tahir, and Z. Liqiang, "Short term
[7] G. Dudek, P. Pełka, and S. Smyl, "A hybrid residual dilated LSTM end electricity load forecasting using hybrid prophet-LSTM model optimized
exponential smoothing model for mid-term electric load forecasting," by BPNN," Energy Reports, vol. 8, pp. 1678-1686, 2022.
arXiv preprint arXiv:2004.00508, 2020. [18] X. Zhou, S. Yang, and S. Sun, "A Deep Learning model for day-ahead
[8] G. E. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung, Time series load forecasting taking advantage of expert knowledge," in 2021 IEEE
analysis: forecasting and control. John Wiley & Sons, 2015. 4th International Electrical and Energy Conference (CIEEC), 2021: IEEE,
pp. 1-5.
[9] M. H. Jifri, E. E. Hassan, and N. H. Miswan, "Forecasting performance
of time series and regression in modeling electricity load demand," in [19] J. Zhang, Y.-M. Wei, D. Li, Z. Tan, and J. Zhou, "Short term electricity
2017 7th IEEE international conference on system engineering and load forecasting using a hybrid model," Energy, vol. 158, pp. 774-781,
technology (ICSET), 2017: IEEE, pp. 12-16. 2018.
[10] K. Zor, O. Timur, and A. Teke, "A state-of-the-art review of artificial [20] V. Dehalwar, A. Kalam, M. L. Kolhe, and A. Zayegh, "Electricity load
intelligence techniques for short-term electric load forecasting," in 2017 forecasting for Urban area using weather forecast information," in 2016
6th international youth conference on energy (IYCE), 2017: IEEE, pp. 1- IEEE International Conference on Power and Renewable Energy
7. (ICPRE), 2016: IEEE, pp. 355-359.
[11] M. Q. Raza and A. Khosravi, "A review on artificial intelligence based [21] A. G. Abdullah, G. M. Suranegara, and D. L. Hakim, "Hybrid PSO-ANN
load demand forecasting techniques for smart grid and buildings," application for improved accuracy of short term load forecasting,"
Renewable and Sustainable Energy Reviews, vol. 50, pp. 1352-1372, WSEAS Transactions on Power Systems, vol. 9, no. 446, p. 51, 2014.
2015. [22] M. R. Haq and Z. Ni, "A new hybrid model for short-term electricity load
[12] methods," Central European Journal of Operations Research, vol. 27, no. forecasting," IEEE access, vol. 7, pp. 125413-125423, 2019.I. S. Jacobs
4, pp. 1033-1049, 2019. and C. P. Bean, “Fine particles, thin films and exchange anisotropy,” in
Magnetism, vol. III, G. T. Rado and H. Suhl, Eds. New York: Academic,
[13] [2] S. R. Salkuti, "Short-term electrical load forecasting using hybrid 1963, pp. 271–350.
ANN–DE and wavelet transforms approach," Electrical Engineering, vol.
100, no. 4, pp. 2755-2763, 2018.

88
Authorized licensed use limited to: Indian Institute of Technology Gandhinagar. Downloaded on August 28,2024 at 14:09:16 UTC from IEEE Xplore. Restrictions apply.