You are on page 1of 4

International Journal of Scientific & Engineering Research Volume 11, Issue 11, November-2020 807

ISSN 2229-5518

Utilization of Double Exponential Smoothing in


Energy Consumption Forecasting for Puerto
Princesa Distribution System
Michael Ernie F. Rodriguez, Alfred Rey G. Vasquez, Roy C. Dayupay

Abstract— Electric load demand forecasting plays a vital role in securing the efficient operation of power systems. There is a need for
energy providers to track electricity periodically if it is adequate. If they don’t, other alternatives should be handled, such as a new power
station, which can provide all of them with ample supplies without interruptions. The researchers believe that evaluating the electricity
demand and reliable forecasts can help the power industry supply its energy and eventually draw investors to create electricity. This will no
longer trigger rotating brownouts and energy shortages in the region. This paper aims to provide accurate load forecasting from 2019 to
2028 in the Puerto Princesa distribution system utilizing double exponential smoothing. Exponential smoothing methods are prediction
approaches that are frequently used as automated forecasts to evaluate univariate time series because of their simplicity and robustness.
The researchers conducted various tests considering different values of smoothing parameters to determine the forecast with the minimum
error. The results obtained demonstrate that the energy consumption can be correctly estimated. Using the smoothing parameters α = 0.1
and β = 0.1, the energy consumption by 2028 is expected to be 393,572.6 MWh with a mean average percent error of 1.29%.

Index Terms— distribution system, energy consumption forecasting, exponential smoothing, load demand, long-term forecast, power
shortage, smoothing parameters

——————————  ——————————

IJSER
1 INTRODUCTION

E lectricity is one of the fundamental needs and essential


resources in sustaining life that people utilize every day.
The electricity demand in the world is anticipated to grow
predictions.
Exponential smoothing is one of the most widely used
mathematical techniques for forecasting. It uses time-series
due to the dependence on electricity of humanity to perform data, which is a collection of organized measurements on the
different tasks. To prepare for the electricity demand growth, quantitative characteristics of a phenomenon at evenly spaced
load forecasting is conducted to estimate the future electricity time points. Double exponential smoothing is a well-known
demand. This study is focused on the Philippine forecasting method, especially in economic and business-
Government's repeated problem of electricity output and related applications [3]. It has been shown that the method of
supply throughout the region. In recent years, several remote time series analysis of the exponential smoothing method is a
provinces have been hit by a power outage in the region. Even kind of excellent efficiency, good adaptability, and a method
the same thing happens in Metro Manila. Many firms and of data analysis [4]. This approach is widely used in stock
sectors are also positively impacted. market estimation, tourism, economic and population logistic
Quaiyum et al. [1] state that precise load forecasting helps analysis, traffic flow, and other natural sciences.
make unit contribution decisions, minimize spinning reserve The single exponential smoothing (SES) approach is a
power, and improve system maintenance schedule. In technique that continually improves estimation by combining
addition to playing a key role in cost control, the efficiency of past values of time series data to reduce the prevalence
power systems is also important. Experts can foresee the (exponential). The SES method, including quantitative
burdens in the future and can also make important choices for prediction methods with historical data patterns, is unreliable
the scheme. There are three types of load forecasting: short- and time series dependent. The characteristics of this approach
term, medium-term, and long-term forecasts, according to [2]. is that the data evaluated are time series and suitable for a
The short-term forecast is used for hourly and weekly horizontal pattern and utilize various parameters for past
predictions, the medium-term forecast is for monthly data. The metric declines exponentially from the magnitude of
predictions, and the long-term forecast is for yearly the most current measurements to the longest measured
values. SES approach is best suited for forecasting variability
————————————————
randomly (irregularly) [5].
 Michael Ernie F. Rodriguez is currently pursuing Master of Science in
Electrical Engineering at Mapua University, Philippines and is currently
The double exponential method of smoothing, also known
an Instructor at Palawan State University, Philippines. E-mail: as Holt’s exponential method for smoothing, is primarily used
mefrodriguez@psu.palawan.edu.ph to provide economies and helps forecasts and patterns to be
 Alfred Rey G. Vasquez is a graduate of Master of Science in Electrical predicted [3], [4], [6]. The Double Exponential Smoothing
Engineering at Mapua University, Philippines and is currently an
Instructor at Palawan State University, Philippines. E-mail: (DES) Holt system has been commonly used for numerous
arvasquez@psu.palawan.edu.ph predictions. This forecasting approach smoothes out real data
 Roy C. Dayupay is a graduate of Master of Engineering at Mapua and patterns with two distinct criteria; therefore, this method
University and is currently an Instructor at Palawan State University,
is ideal for projecting data for a particular trend. The DES
Philippines. E-mail: rdayupay@psu.palawan.edu.ph
approach is more effective than other approaches for
IJSER © 2020
http://www.ijser.org
International Journal of Scientific & Engineering Research Volume 11, Issue 11, November-2020 808
ISSN 2229-5518

modeling data patterns. The DES approach ensures high 2.2 Forecasting Using Double Exponential Smoothing
forecasting accuracy of the data model by considering patterns The double exponential smoothing is used when the data
that exist. DES is more reliable than Linear Pattern, Quadratic shows a trend. The level and trend are the two components
Trend, Exponential Rise, Moving Average (MA), Auto updated in each period [9]. The following is the formula used
Regressive Integrated Moving Average (ARIMA), and Kalman for double exponential smoothing [3], [4]:
Filter methods [7], [8]. St  Yt  (1   )(St 1  Tt 1 ) (1)
In this paper, the Puerto Princesa Distribution System's
energy consumption for the year 2019-2028 was forecasted Tt   (St  St 1 )  (1   )Tt 1 (2)
using double exponential smoothing.
Yˆt  n  St  nTt (3)
where St and Tt are the smoothed level and trend,  and 
2 METHODOLOGY are smoothing parameters (0 <  < 1, and 0 <  < 1), Yt is the
This part of the paper discusses the procedure in conducting actual value of the data in period t, and Yˆt  n is the value of the
this study. Fig. 1 shows the conceptual framework of the forecast done from the origin of the forecast t.
study. There are several methods for determining the initial values
of the smoothed level and trend [9]. In this paper, linear
regression is used between the period and the energy
consumption to determine the initial values S1 and T1.
S1  Y1  (1   )(S0  T0 ) (4)
T1   (S1  S0 )  (1   )T0 (5)
The model equation for linear regression and the formula for
solving the values of the slope and intercept are:

IJSER
Y  a  bX (6)

n
( Xt  X)(Yt  Y )
b t 1
(7)

n
t 1
( Xt  X)2
a  Y  bX (8)
where Xt is the period t, Yt is the energy consumption at time
t, Y and X are the mean of Y and X, respectively, a is the
intercept, and b is the slope. From (4) and (5), the value of S0 is
the intercept, and the value of T0 is the slope [8].

2.3 Measuring Forecast Error


The measurement of the forecast error is essential in
forecasting to determine the accuracy of the forecasting model.
The most commonly used measuring tools for measuring the
error rate of a forecasting model are the mean absolute error
(MAE), mean square error (MSE), and mean absolute
percentage error (MAPE) [8]. The formulas used for the
calculation of the MAE, MSE, and MAPE are shown in (9),
(10), and (11).

n
t 1
Yt  Yˆt
MAE  (9)
n

Y  Yˆ 
2

n
t 1 t t
MSE  (10)
n
Yt  Yˆt

n
Fig. 1. Conceptual Framework t 1
 100%
Yt
MAPE  (11)
n
2.1 Collection of Historical Data where Yt is the actual value, Yˆt is the forecasted value, and n
The annual energy consumption for 2014-2018 of the Puerto is the number of data points. The values of MAPE for
Princesa Distribution System was collected from the utility evaluating the forecast's accuracy are shown in Table 1 [8].
company. These will be used for the forecast in the next ten
years using double exponential smoothing.
IJSER © 2020
http://www.ijser.org
International Journal of Scientific & Engineering Research Volume 11, Issue 11, November-2020 809
ISSN 2229-5518

TABLE 1 TABLE 4
MAPE VALUES FOR FORECAST EVALUATION FORECAST ERRORS OF VARIOUS SMOOTHING PARAMETERS
Value of MAPE Forecast Accuracy Test   MAE MSE MAPE
Less than 10% High 1 0.1 0.1 2,508.9964 9,577,666.8743 1.29%
Between 10% to 20% Good 2 0.1 0.2 2,532.4473 9,716,529.5716 1.30%
Between 20% to 50% Reasonable 3 0.1 0.3 2,556.0885 9,859,826.6875 1.32%
Greater than 50% Low 4 0.1 0.4 2,579.9057 10,007,456.1252 1.33%
5 0.1 0.5 2,603.8845 10,159,331.4860 1.34%
6 0.2 0.1 2,677.4196 10,905,136.1924 1.38%
3 RESULTS AND DISCUSSIONS 7 0.2 0.2 2,722.8946 11,224,181.5432 1.40%
The historical data of the annual energy consumption for 2014- 8 0.2 0.3 2,768.6298 11,559,408.7349 1.42%
2018, as shown in Table 2, were collected from Palawan 9 0.2 0.4 2,814.5252 11,910,526.5683 1.45%
Electric Cooperative. It was observed that the energy 10 0.3 0.1 2,823.3289 12,333,812.8949 1.45%
consumption in the Puerto Princesa Distribution System is 11 0.2 0.5 2,860.4830 12,277,338.5618 1.47%
increasing. 12 0.3 0.2 2,886.9013 12,867,084.9317 1.48%
13 0.4 0.1 2,944.1685 13,840,352.1714 1.51%
TABLE 2 14 0.3 0.3 2,950.1265 13,433,737.2267 1.52%
HISTORICAL DATA ON ENERGY CONSUMPTION 15 0.3 0.4 3,012.7173 14,032,963.2338 1.55%
Year t Energy Consumption (MWh) 16 0.4 0.2 3,020.3254 14,608,931.6344 1.55%
2014 1 149,258.4740 17 0.5 0.1 3,039.5770 15,399,489.8458 1.56%
2015 2 161,749.1021 18 0.3 0.5 3,074.3986 14,663,999.8312 1.58%
2016 3 180,486.9336 19 0.4 0.3 3,094.5526 15,428,662.4880 1.59%
20 0.5 0.2 3,122.1325 16,406,763.6380 1.60%

IJSER
2017 4 194,850.9336
2018 5 221,079.8156 21 0.4 0.4 3,166.2891 16,296,431.6280 1.63%
22 0.5 0.3 3,200.1624 17,476,402.9934 1.64%
23 0.4 0.5 3,235.0122 17,208,671.0736 1.66%
To determine the initial values of level and trend, linear
24 0.5 0.4 3,272.8049 18,599,560.0984 1.68%
regression was applied using the MS Excel® Analysis
25 0.5 0.5 3,339.2921 19,766,065.0911 1.71%
Toolpak. The linear regression yields an intercept value of
128,461.6974 and a slope value of 17,674.4515. These intercept
and slope values correspond to the values of S0 and T0, Upon comparing MAPE from the tests in Table 4, the
respectively. Using the smoothing parameters  = 0.1 and  lowest MAPE obtained was 1.29% with the smoothing
= 0.1, the smoothed level and trend initial values (S1 and T1) parameters  = 0.1 and  = 0.1. These values of smoothing
are 146,448.3814 and 17,705.6747. The values of the smoothed parameters were used to forecast the energy consumption for
level and trend for t = 1 up to t = 5 is shown in Table 3. the next ten years. The forecast model using double
exponential smoothing at  = 0.1 and  = 0.1 is
TABLE 3 Yˆ  216,824.7899  17,674.78103n
5n (12)
VALUES OF SMOOTHED LEVEL AND TREND FROM 2014-2018
The forecast results are shown in Table 5 and Figure 2, which
Energy Consumption Smoothed indicate that energy consumption through the next ten years
t Trend
(MWh) Level from 2019 to 2028 is increasing.
1 149,258.4740 146,448.3814 17,705.67472
2 161,749.1021 163,913.5607 17,681.62518 TABLE 5
3 180,486.9336 181,484.3606 17,670.54266 FORECASTED ENERGY CONSUMPTION FOR THE YEAR 2019-2028
4 194,850.9336 198,724.5063 17,627.50296 Year n Energy Consumption (MWh)
5 221,079.8156 216,824.7899 17,674.78103 2019 1 234,499.57
2020 2 252,174.35
The mean absolute error (MAE), mean square error (MSE), 2021 3 269,849.13
and mean absolute percentage error (MAPE) at  = 0.1 and 2022 4 287,523.91
 = 0.1 was evaluated. The results obtained are MAE = 2023 5 305,198.70
2,508.9964, MSE = 9,577,666.8743 and MAPE = 1.29% 2024 6 322,873.48
The researchers also tested other values of the smoothing 2025 7 340,548.26
parameters to be used in the forecast. The results for these 2026 8 358,223.04
tests are shown in Table 4, along with the forecast errors MAE, 2027 9 375,897.82
MSE, and MAPE. 2028 10 393,572.60

IJSER © 2020
http://www.ijser.org
International Journal of Scientific & Engineering Research Volume 11, Issue 11, November-2020 810
ISSN 2229-5518

cation,” 2012 Int. Conf. on Advanced Mechatron. Syst. ICAMechS 2012,


pp. 386–388, 2012.
[7] Muladi, S. A. Siregar, and A. P. Wibawa, “Double Exponential-
Smoothing Neural Network for Foreign Exchange Rate Forecasting,”
Proc. - 2nd East Indones. Conf. Comput. Inf. Technol. Internet Things Ind.
EIConCIT 2018, pp. 118–122, 2018, doi:
10.1109/EIConCIT.2018.8878591.
[8] R. Anggrainingsih, A. Prabanuadhi, and S. P. Yohanes, “Forecasting
the Number of Patients at RSUD Sukoharjo Using Double Exponen-
tial Smoothing Holt,” Proceeding - 2018 Int. Conf. ICT Rural Dev. Rural
Dev. through ICT Concept, Des. Implic. IC-ICTRuDEv 2018, pp. 54–58,
2018, doi: 10.1109/ICICTR.2018.8706850.
Fig. 2. Historical and Forecasted Energy Consumption from the Year [9] P. Kalekar, “Time series forecasting using Holt-Winters exponential
2014-2028 smoothing,” Kanwal Rekhi Sch. Inf. Technol., no. 04329008, pp. 1–13,
2004, [Online]. Available:
http://www.it.iitb.ac.in/~praj/acads/seminar/04329008_Exponenti
alSmoothing.pdf.
4 CONCLUSION
In this study, double exponential smoothing was utilized to
forecast the Puerto Princesa Distribution System's energy
consumption from 2019 to 2028. The researchers conducted
various tests to determine the forecast with the minimum

IJSER
error by varying the smoothing parameters  and . The
smoothing parameters with a value  = 0.1 and  = 0.1
obtained the lowest MAPE of 1.29% based on the tests
conducted. Using these smoothing parameters, the energy
consumption in the year 2028 is expected to be 393,572.6
MWh.
For future works, the double exponential smoothing can be
applied to other locations and other distribution systems. The
researchers also recommend applying optimization techniques
to determine the smoothing parameters which can produce
minimum forecast error.

REFERENCES
[1] S. Quaiyum, Y. I. Khan, S. Rahman, and P. Barman, “Artificial Neural
Network based Short Term Load Forecasting of Power System,” Int.
J. Comput. Appl., vol. 30, no. 4, pp. 975–8887, 2011.
[2] M. Y. Khamaira, A. S. Krzma, and A. M. Alnass, “Long Term Peak
Load Forecasting For The Libyan Network,” no. December, pp. 185–
193, 2018, doi: 10.21467/proceedings.2.23.
[3] A. C. Adamuthe, R. A. Gage, and G. T. Thampi, “Forecasting cloud
computing using double exponential smoothing methods,” ICACCS
2015 - Proc. 2nd Int. Conf. Adv. Comput. Commun. Syst., pp. 3–7, 2015,
doi: 10.1109/ICACCS.2015.7324108.
[4] A. Chusyairi, N. S. P. Ramadar, and Bagio, “The use of exponential
smoothing method to predict missing service e-report,” Proc. - 2017
2nd Int. Conf. Inf. Technol. Inf. Syst. Electr. Eng. ICITISEE 2017, vol.
2018-Janua, no. February, pp. 39–44, 2018, doi:
10.1109/ICITISEE.2017.8285535.
[5] R. Gustriansyah, N. Suhandi, F. Antony, and A. Sanmorino, “Single
exponential smoothing method to predict sales multiple products,” J.
Phys. Conf. Ser., vol. 1175, no. 1, 2019, doi: 10.1088/1742-
6596/1175/1/012036.
[6] F. Chen, Q. Li, J. Liu, and J. Zhang, “Variable smoothing parameter
of the double exponential smoothing forecasting model and its appli-

IJSER © 2020
http://www.ijser.org

You might also like