Uploaded by Baldemar Aguirre Fraire

Renewable Energy Forecasting Model: A Case Study of Korea

advertisement
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Contents lists available at ScienceDirect
Renewable and Sustainable Energy Reviews
journal homepage: http://www.elsevier.com/locate/rser
A deep learning-based forecasting model for renewable energy scenarios to
guide sustainable energy policy: A case study of Korea
KiJeon Nam a, 1, Soonho Hwangbo b, 1, ChangKyoo Yoo a, *
a
Dept. of Environmental Science and Engineering, College of Engineering, Center for Environmental Studies, Kyung Hee University, Seocheon-dong 1, Giheung-gu, YonginSi, Gyeonggi-Do, 446-701, South Korea
b
Process and Systems Engineering Center (PROSYS), Department of Chemical and Biochemical Engineering, Technical University of Denmark, Søltofts Plads 229, 2800,
Kgs. Lyngby, Denmark
A R T I C L E I N F O
A B S T R A C T
Keywords:
Renewable energy forecasting
Deep learning
Sustainable energy policy
Renewable energy scenario
Techno-economic-environmental analysis
Jeju island
Renewable and sustainable energy systems and policies have globally been promoted to transition from fossil fuel
sources to environmentally friendly renewable energy sources such as wind power, photovoltaic energy, and fuel
cells. Wind and solar energy sources are erratic and difficult to implement in renewable energy systems,
therefore, circumspection is needed to implement such renewable energy systems and policies. Accordingly, this
study develops an energy forecasting model with renewable energy technologies on which policy can be based,
using the Korean energy policy as a case study. Deep learning-based models forecast fluctuating variation in
electricity demand and generation, which are necessary in renewable energy system but not possible with
conventional models. The gated recurrent unit shows the best prediction performance among the forecasting
models evaluated, and is therefore selected as the base model to evaluate four different renewable energy sce­
narios. The scenarios are evaluated according to economic-environmental cost assessment. The optimal scenario
uses an integrated gasification combined cycle, onshore and offshore wind farms, photovoltaic power stations,
and fuel cell plants; in particular, this scenario shows the lowest economic-environmental costs, generates stable
electricity for demand, and achieves a policy with 100% renewable energy. The optimal scenario is assessed by
considering its strengths, weaknesses, opportunities, and threats analysis while also considering technoeconomic-environmental domestic and global energy circumstances.
1. Introduction
Sustainable development and global climate change are important
issues in the 21st century with regard to energy. Energy consumption is
increasing by 2% per year, and overall energy generation is currently
dependent on fossil fuels [1]. Anthropogenic greenhouse gas (GHG)
emission due to the usage of fossil fuels, is growing considerably,
causing abnormal climate across the globe, including droughts and
heavy rainfall [2,3]. Moreover, it is estimated that GHG will increase by
30% in 20 years without any regulatory restriction on the use of fossil
fuels [1]. Therefore, global efforts regarding renewable and sustainable
energy systems and policies, including energy management, have been
conducted to overcome the dependency of fossil fuel power plants and
mitigate global climate change and negative environmental effects due
to fossil fuels [4,5].
The Korean energy system is highly dependent on fossil fuels such as
coal, oil, and liquefied natural gas, which make up 64.4% of its energy
sources. Thus, Korea was ranked first with regard to increased GHG
emission rates from 1990 to 2014, and sixth regarding GHG emissions by
the Organization for Economic Co-operation and Development in 2015
[6,7]. As a result of the energy problem of Korea, the Korean govern­
ment has strived to reduce GHG emissions while increasing the use of
renewable energy facilities and developing renewable energy systems.
One of the more remarkable renewable energy systems of Korea is the
use of 100% renewable energy on Jeju Island. Jeju island has been
designated as a 100% renewable energy system island with the imple­
mentation of wind farms, photovoltaic power stations, and fuel cell
plants [7].
Several studies investigated the feasibility of the policy and sug­
gested renewable energy systems on Jeju island. Kim et al. [8] suggested
a sustainable energy management system composed of a wind turbine, a
* Corresponding author.
E-mail address: ckyoo@khu.ac.kr (C. Yoo).
1
The first and second authors have identical collaboration in this paper.
https://doi.org/10.1016/j.rser.2020.109725
Received 26 March 2019; Received in revised form 30 December 2019; Accepted 21 January 2020
Available online 4 February 2020
1364-0321/© 2020 Elsevier Ltd. All rights reserved.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
K. Nam et al.
Abbreviations
ANN
ARIMA
ARMA
CCS
DNN
EMD
GHG
GRU
HSS
HVDC
IGCC
IMFs
LSTM
MAE
MASE
ML
MLR
ReLU
SARIMA
SWOT
artificial neural network
autoregressive integrated moving average
auto regressive moving average
carbon capture and sequestration
deep neural networks
empirical mode decomposition
greenhouse gas
gated recurrent unit
hydrogen storage system
high voltage direct current
integrated gasification combined cycle
intrinsic mode functions
long short-term memory
mean absolute error
mean absolute scaled error
machine learning
multiple linear regression
rectified linear unit
seasonal autoregressive integrated moving average
strengths, weaknesses, opportunities, and threats
Nomenclature
Crenewable energy facility capacity of renewable energy facility (MW)
Nrenewable energy facility number of renewable energy facilities
$renewable energy facility economic and environmental cost of renewable
energy facility ($/year)
photovoltaic panel, a battery, and an electricity converter. The hybrid
system was evaluated in terms of the net present cost and the cost of
electricity. Kwon et al. [9] proposed a renewable energy supply system
using the wind turbine, the photovoltaic panel, and the battery. They
also evaluated the system considering variations in equipment costs,
wind, and solar radiation. Although these studies presented notable
results, they were conducted based on a monthly measured data set, and
only evaluated the operation costs. Renewable energy-related informa­
tion such as daily complex patterns of energy demand and supply,
electricity generation facility costs considering the economy and envi­
ronment, and the selection of renewable energy facilities with their
capacities can affect the economic and environmental feasibility of
renewable energy systems [5,10,11]. Therefore, it is essential to
consider the complex patterns of the renewable energy through a fore­
casting model and relate these patterns with the sustainable energy
system and policy; this is because the forecasting model reflects erratic
and uncertain characteristics of the energy demand and supply [11,12].
A renewable energy forecasting model with a long-term time scale
(seven-days ahead) can be utilized for a feasibility study of the energy
system design policy, and moreover, this model can reduce unnecessary
regulatory costs while implementing renewable energy sources into the
energy system [11,13,14].
We developed a forecasting model with seven-days ahead electricity
demand and renewable energy generation based on deep learning
techniques and domain knowledge. These results can be used to promote
and guide feasible renewable energy systems and policies. This study
compares and evaluates deep learning models and conventional statis­
tical models. The deep learning models include deep neural networks
(DNN), long short-term memory (LSTM), and gated recurrent unit
(GRU), and overcome the disadvantages of conventional statistical
models such as multiple linear regression (MLR) and seasonal autore­
gressive integrated moving average (SARIMA). Comparison and evalu­
ation of the forecasting models are significant since deep learning
models can have different performances depending on the properties of
the data. The performances of deep learning models differ according to
the forecasting time, training duration, target data, and simple or
ensemble structure of the models [15,16]. However, there are several
issues to consider when selecting an appropriate forecasting model.
Thus, we thoroughly compare and evaluate the forecasting models using
accurate numerical evaluators and confidence intervals, and select the
best forecasting model for future electricity demand and renewable
energy generation. We then utilize the proposed model for renewable
energy scenarios for the policy design of Jeju Island to achieve their
energy policy (see Section 3.1 and Appendix A). The renewable energy
sources considered in this study are wind power, photovoltaic power,
fuel cells, an integrated gasification combined cycle (IGCC), and a
hydrogen storage system (HSS). In this study, the following four sce­
narios are compared: (1) wind power, photovoltaic power, and fuel cells,
(2) IGCC, wind power, photovoltaic power, and fuel cells, (3) wind
power, photovoltaic power, fuel cells, and HSS, and (4) IGCC, wind
power, photovoltaic power, fuel cells, and HSS. The renewable energy
capacity for each scenario is determined by considering the electricity
demand of the target region. Then, the scenarios are evaluated by eco­
nomic and environmental aspects; the cheapest renewable energy sce­
nario that produces 100% energy is assessed according to its strengths,
weaknesses, opportunities, and threats (SWOT) considering
techno-economic-environmental domestic and global renewable energy
circumstances. The evaluated scenarios can guide the policy makers and
decision-makers of Jeju Island.
2. Literature review
2.1. Renewable energy forecasting models
Historical studies related to renewable energy, as summarized in
Table 1, have been conducted. First is the forecasting of electricity de­
mand and renewable energy generation. The forecasting of electricity
demand and renewable energy generation can provide an appropriate
basis to manage and plan the energy supply and design policy [12,14].
Several studies have developed time-series statistical energy forecasting
models for stable energy infra-structures to establish an exact energy
plan. For example, Thatcher [17] predicted a future local electricity
demand curve for Australia using a multiple linear regression (MLR)
model that used local past electricity demand and weather conditions.
Bianco et al. [18] presented a multiple regress model to predict annual
energy consumption up to 2030 using an annual gross domestic product
and expected population in Italy. Goia et al. [19] forecasted the
short-term energy demand caused by heating using a functional linear
regression model. Barak and Sadegh [20] suggested a hybrid autore­
gressive integrated moving average (ARIMA)-adaptive neuro-fuzzy
inference system to forecast the annual electricity consumption in
Table 1
A historical review of renewable energy; energy forecasting and renewable en­
ergy policies.
Criterion
Description
References
Energy
forecasting
Energy demand forecasting using a conventional
time-series forecasting model
An ARIMA-based forecasting model for renewable
energy generation
Electricity demand forecasting model using a
neural network
Ensemble machine learning models to forecast
renewable energy
Current and future energy policy assessment
Renewable energy scenario-based future energy
analysis
[17–20]
Political
strategies
2
[21–25]
[27–30]
[31–35]
[38–42]
[43–46]
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Iran. Reikard [21] forecasted solar irradiance using the ARIMA model,
which considered the log value of the observed solar irradiance.
Palomares-Salas et al. [22] compared the performance of ARIMA and
neural network model to forecast short-time intervals wind speed, and
the ARIMA showed better performance. Li et al. [23] used the auto
regressive moving average (ARMA) and wavelet transform to predict the
wind speed for instituting a wind power system plan. Hejase and Assi
[24] suggested a times-series regression with Box-Jenkins ARMA model
for predicting solar radiation. Vagropoulos et al. [25] presented the
seasonal autoregressive integrated moving average (SARIMA) and
SARIMA model with an exogenous factor to forecast the energy gener­
ation of grid-connected photovoltaic plants. However, it is difficult to
assure the forecasting accuracy of these statistical models because each
model is very complex, structure selection is not easy, and prediction
accuracy decreases with increasing prediction time. Currently, statistical
forecasting models are unreliable, making them inappropriate as a factor
in decision making [13,26].
Thus, the machine learning (ML) forecasting model, which has the
ability to interpret the relationship between the input and output data
without an explicit algorithm, has been used to overcome the disad­
vantages of conventional statistical forecasting models. Geem and Roper
[27] proposed an artificial neural network (ANN) model to predict the
overall energy demand of Korea, and this model considered the gross
domestic product, population, import, and export amounts as input
variables. Ekonomou [28] used an ANN to forecast the long-term energy
consumption of Greece, and the ANN model was compared with a
conventional model to check the superiority of the model. Kaytez et al.
[29] developed least squares support vector machines to forecast the
energy consumption of Turkey. They compared the performance of least
squares support vector machines and the ANN model considering the
specificity and sensitivity of a receiver operating characteristic analysis.
Muralitharan et al. [30] suggested a neural network-based genetic al­
gorithm and a neural network-based particle swarm optimization model
to forecast the electricity demand of a smart grid composed of 750
houses. Li and Shi [31] evaluated three ML forecasting models (feed
forward back propagation, radial basis function, and adaptive linear
element ANN) to predict the wind speed, emphasizing a method to
combine forecasts from different ANN models. Chang [32] applied the
radial basis function ANN to forecast the wind power for a wind energy
conversion system. Troncoso et al. [33] evaluated the model perfor­
mances when predicting short-term wind speed forecasting. Ensemble
models that implemented a k-nearest neighbors with local non-linear
model and weighted k-nearest neighbors with local non-linear model
showed the best performance among the eight regression trees-based
models. Galaet al. [34] proposed a hybrid machine learning model
that uses support vector regression, gradient boosted regression, and
random forest regression to predict solar radiation directly related to
solar energy generation. Heinermann and Kramer [35] forecasted wind
power production by suggesting a heterogeneous ensemble ML model
composed of a decision tree and support vector machine.
However, the forecasting performance of a normal ML method is
limited when using it by itself. However, this performance can be
improved by combining it with other methods. In this context, deep
learning is a branch of ML that models high-level abstractions and has
the ability to describe a hidden high-level invariant structure and
inherent features from the data. Deep learning can also be used as an
effective alternative technique to overcome the limitations of conven­
tional statistical forecasting models and simple ML forecasting models.
Despite these advantages of deep learning, research in this area is recent,
and applications of such deep learning techniques to energy forecasting
is still in the early stages [36,37].
several variables. Kim and Park [38] presented a politically motivated
smart grid system in Korea considering the overall energy, energy con­
sumers, transportation, sustainable energy, and electricity services by
separating the model into three future stages. Aslani et al. [39] sug­
gested an energy model evaluating Finnish renewable energy policies by
analyzing the energy policy scenarios and dependency on imported
energy. Scarlat et al. [40] analyzed a renewable energy policy of the
European Union and determined the major energy sources that will in­
crease the use of renewable energy. He et al. [41] proposed energy
policy path stages for overcoming the limitations of the Chinese sus­
tainable energy policy. Han and Baek [42] compared the capacity of
global renewable energy generation with that of Korea, and suggested
future political strategies and directions to increase the use of renewable
energy. Mathiesen et al. [43] suggested smart renewable energy systems
considering several energy scenarios, including gasified biomass, gas
storage, electric vehicles, electro-fuel production, and heating systems.
Cho and Kim [44] analyzed the Korean electricity supply system sce­
narios for future renewable energy sources from economic, energy se­
curity, and environmental viewpoints. Tripathi et al. [45] found a way
to provide electricity to India using renewable energy based on the
current renewable energy system of India. Connolly et al. [46] analyzed
smart energy system scenarios for European renewable energy systems
considering the energy, environment, and economics. However, in all of
these analyses, the predicted renewable energy supply and total elec­
tricity demand of the countries were assumed to be non-dynamic. En­
ergy systems are complex because electricity demand and generation
have intricate physical and social interactions. Therefore, both energy
planning and the policy need to be facilitated through the use of
deep-exact forecasting models that can be linked to the energy, econ­
omy, and environment while accurately predicting the dynamic elec­
tricity demand and generation [9,12,14,47].
3. Energy forecasting and renewable energy scenarios
The proposed framework of this study for forecasting electricity
demand and renewable energy generation and analyzing renewable
energy scenarios is graphically shown in Fig. 1. The first stage is
implementation of a forecasting model to predict variations in electricity
demand and renewable energy generation. Electricity demand and
renewable energy generation via wind power and photovoltaic power
were collected from 2013 to 2017 on Jeju Island and are shown in Figs. 1
(a) and Fig. 2. The demanded and supplied electricity are measured in
total megawatts per day. Then, the data are decomposed through
empirical mode decomposition (EMD), as shown in Fig. 1 (b), to enhance
the performances of the forecasting models by constructing intrinsic
mode functions (IMFs) from the nonlinear time-series data. Fig. 1 (c)
shows an algorithm used in the forecasting model. The decomposed data
are input into the forecasting models, and the original electricity data,
which are not decomposed, are forecasted by the model. The forecasting
models use a moving window method that utilizes 21 days of data for
learning and forecasting electricity data for the next seven days. This
window moving technique has an advantage in that it enhances the
time-dependent forecasting performance of the models. As shown in
Fig. 2, four years of data, from 2013 to 2016, are used as the training set,
and data for 2017 are used to test the forecasting models. This study uses
two statistic forecasting models—MLR and seasonal autoregressive in­
tegrated moving average (SARIMA)—and three deep learning-based
forecasting models—GRU, LSTM, and DNN. The electricity demand
and the generated renewable electricity have highly fluctuating char­
acteristics, which can be issues for forecasting future electricity demand
and generation using only the forecasting model. Therefore, it is crucial
to extract hidden patterns in the data by utilizing domain knowledge
such as the EMD. The extracted domain knowledge not only reduces
irrelevant information in forecasting algorithms, but also improves the
forecasting accuracy [48,49]. Therefore, decomposed data for the pre­
vious 21 days are employed as the domain knowledge to forecast
2.2. Political strategies
Several studies have analyzed and proposed current and future en­
ergy scenarios to achieve a renewable energy policy by considering
3
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Fig. 1. Graphical diagram of the forecasting model and renewable energy scenario analysis; (a)–(g) indicate sub-stages of the proposed framework.
electricity for the next seven days. Fig. 1 (d) shows that the forecasted
electricity data model is validated compared to the observed data set
with respect to the mean absolute scaled error (MASE) and mean ab­
solute error (MAE). The MASE is suitable for measuring the forecast
accuracy when the data have a varied scale, and the MAE is generally
used as a model evaluator, measuring differences between the forecasted
and observed data [50,51]. The MASE and MAE are defined as:
0
1
MASE ¼
T B
B
1X
B
T t¼1 B
@
1
T 1
T
P
MAE ¼ t¼1
jyt
T
b
yt j
C
C
jyt by t j
C
C
T
P
jyt yt 1 jA
y t are the observed data and forecasted data at time t,
where yt and b
respectively.
The second stage is the suggestion and analysis of a renewable en­
ergy scenario on Jeju Island. Based on the quantities of forecasted
electricity demand and renewable energy generation in 2017, four
renewable energy scenarios are suggested to supply electricity using
100% renewable energy technologies such as wind power, photovoltaic
power, fuel cells, and IGCC. The suggested renewable energy scenario is
graphically described in Fig. 1 (e). The different scenarios, which are
explained in Section 3.4 in detail, individually suggest a feasible com­
bination of the renewable energy facilities’ capacities to satisfy timevarying electricity demands with the least cost. Economic costs in
terms of construction, operation and management, and environmental
costs incurred due to the emission of pollutants from energy facilities are
the analysis factors considered for the scenarios, as shown in Fig. 1 (f).
The scenario that has the least total economic-environmental costs
(1)
t¼2
(2)
4
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
considering the global status of the energy system and market.
3.1. Current energy system status and energy policy of Jeju Island
The current energy generation system of Jeju Island mainly consists
of fossil fuels instead of renewable energy. Therefore, deep insights and
considerations are needed to achieve 100% electricity generation by
renewable energy technologies in the future. This study suggests an
accurate forecasting model to design an acceptable renewable and sus­
tainable energy policy corresponding to Jeju Island’s characteristics.
The details of the current energy system and energy policy of Jeju Island
are discussed in Appendix A.
3.2. Forecasting model for electricity demand and renewable energy
generation
The renewable energy generation and total electricity demand of
Jeju Island are predicted by forecasting models that combine data
decomposition techniques and conventional statistical and deep
learning prediction techniques. This combination of techniques results
in a high performance for feature extracting and forecasting. The
decomposition algorithm improves the forecasting ability of the model
by providing decomposed flexible sub data from intermittent raw data.
3.2.1. Empirical mode decomposition (EMD)
Empirical mode decomposition (EMD) is a widely used decomposi­
tion algorithm. EMD is fully data-driven, adaptive, and does not need
predetermined transforms that depend on the selection of a specific
structure. EMD decomposes non-linear and non-stationary data xðtÞ into
intrinsic mode functions (IMFs) satisfying two conditions: the number of
extrema and zero-crossings are equal or have a difference of one, and the
mean value of the envelopes defined by the local maxima and minima is
zero or close to zero. As shown in Fig. 1 (b), this study uses EMD to
enhance the performance of the forecasting model, and the computed
IMFs and residue are input into the forecasting model. The main algo­
rithm of EMD is described in Appendix B with Eq. (B.1).
3.2.2. Conventional statistical forecasting models
The sub-series decomposed data by EMD is used as the input for the
statistical and deep learning forecasting models. This study uses con­
ventional statistical forecasting models MLR, SARIMA, deep learning
models DNN, LSTM, and GRU, as shown in Fig. 1 (c). The MLR predicts a
dependent variable y using two or more explanatory independent vari­
ables x via linear equation fitting [52]. Additionally, SARIMA is the
extended ARIMA model and improves the forecasting ability by
removing seasonal variation trends by considering the differences. The
details of MLR and SARIMA are described in Appendix C.
3.2.3. Deep learning-based forecasting models
Deep learning models are recent machine learning methods, but have
rarely been implemented in the field of energy forecasting. In this study,
such models are used with EMD for energy forecasting to overcome the
weaknesses of conventional statistical forecasting models. Among the
several deep learning models, this study utilizes DNN as a fundamental
deep learning model and LSTM and GRU to consider the time-series
characteristics of the data. DNN uses a multi-layer structure consisting
of one input layer, one output layer, and several hidden layers based on a
hierarchical model. The rectified linear unit (ReLU) and dropout are
used to overcome the vanishing gradient problem and over-fitting. The
ReLU evades gradient vanishing by converting negative inputs to zero,
and the dropout randomly drops units from the hidden layers while
computing activations in the forward pass and updating weights in the
backward pass [53].
LSTM and GRU are the improved recent versions of conventional
deep recurrent neural networks (deep RNN). Deep RNN has three major
disadvantages: training takes a long time, deep RNN is not able to learn
Fig. 2. Electricity data of Jeju island from 2013 to 2017; (a) electricity de­
mand, (b) wind electricity generation, and (c) photovoltaic elec­
tricity generation.
among the scenarios is selected as the optimal scenario to provide
guidance for sustainable energy policy design. Finally, the optimal
scenario is assessed using a SWOT analysis to overcome various weak­
nesses and threats, as well as maximize the strengths and opportunities.
The techno-economic-environmental SWOT analysis includes domesticinternal and global-external assessments, as shown in Fig. 1 (g). The
internal assessment illustrates the strengths and weaknesses of the
optimal scenario considering the energy circumstances of Korea, and the
external assessment demonstrates the opportunities and threats by
5
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
suggested. The information from the renewable energy facilities
is summarized in Table F1. The environmental costs, shown in
Table F2, are the environmental marginal damage costs consid­
ering climate change damage [61–63]. All costs are converted to
the 2017 equivalent dollar using the consumer price index.
long-term patterns, and information related to the initial data gradually
fades away. Therefore, LSTM and GRU were developed to solve these
problems by implementing deep and long-term memory cells in their
structures, similar to their applications in natural language processing
[54]. Despite these advantages of deep learning, their use in energy
forecasting is recent and studies related to energy forecasting are still in
the early stages [36]. The details of LSTM and GRU are provided in
Appendix D.
The number of renewable energy facilities in each scenario is
calculated by minimizing the economic and environmental total costs.
For instance, the objective function for calculating economic and envi­
ronmental costs in the case of scenario 1 is expressed by Eq. (3).
3.3. Renewable energy technologies
f ¼ Nonshore wind farm � $onshore wind farm þ Noffshore wind farm � $offshore wind farm
þNphotovoltaic power station � $photovoltaic power station þ Nfuel cell plants � $fuel cell plants
Based on the proposed policy that does not utilize fossil fuels or other
methods resulting in the emission of carbon dioxide, this study uses
wind farms, photovoltaic power stations, and fuel cells. Additional
sustainable energy technologies, IGCC and HSS, are used to assess
renewable energy scenarios in this study to supply stable electricity not
affected by variations in the weather or season. Detailed descriptions of
the renewable and sustainable energy technologies are provided in
Appendix E.
Nonshore wind farm � Conshore wind farm þ Noffshore wind farm � Coffshore wind famr
� 8 � Nphotovoltaic power station � Cphotovoltaic power station
(3)
(4)
Nonshore wind farm � Conshore wind farm þ Noffshore wind farm � Coffshore wind farm
þNphotovoltaic power station � Cphotovoltaic power station þ Nfuel cell plants � Cfuel cell plants
� Necessary capacity
3.4. Description of the renewable energy scenarios
(5)
Here, Nrenewable energy facility is the number of renewable energy facilities
(e.g., Nonshore wind farm is the number of onshore wind farm),
$renewable energy facility is the economic and environmental cost of the
renewable energy facility, and Crenewable energy facility is the capacity of the
renewable energy facility. Note that this study uses the nominal capacity
suggested by EIA [60] as summarized in Table F1. Eqs. (4) and (5) are
constraint functions on the capacity determination of the renewable
energy facilities. The first constraint is formulated based on the expan­
sion plan of renewable energy facilities of Jeju Island, where the ca­
pacity of the wind farms is eight times that of the photovoltaic power
stations. The second constraint is that the total capacity of renewable
energy technologies should be higher than the necessary capacity, which
is 120% of the forecasted electricity demand. In addition, the following
boundary conditions of the wind farms are applied:
The four scenarios in this study are introduced to help inform sus­
tainable energy policy design and decisions for Jeju Island (Fig. 1 (e)).
The prepared scenarios are:
(1) Scenario 1: offshore wind farms, onshore wind farms, photovol­
taic power stations, and fuel cell plants; base scenario that is
based on the current sustainable energy plan of Jeju Island.
(2) Scenario 2: offshore wind farms, onshore wind farms, photovol­
taic power stations, fuel cell plants, and IGCC; use of the addi­
tional IGCC technology to supply stable electricity.
(3) Scenario 3: offshore wind farms, onshore wind farms, photovol­
taic power stations, fuel cell plants, and HSS; use of the additional
HSS to store surplus generated electricity and supply the stored
electricity.
(4) Scenario 4: offshore wind farms, onshore wind farms, photovol­
taic power stations, fuel cell plants, IGCC, and HSS; use of addi­
tional sustainable energy technology and an energy storage
system.
0 � Nonshore wind farm � 2
(6)
1 � Noffshore wind farm � 5
(7)
Determining the maximum numbers of onshore and offshore wind
farms is one goal of the policy, and the minimum numbers of onshore
and offshore wind farms are based on the ongoing construction projects
on Jeju Island. The HSS is used in scenarios 3 and 4, and is determined
using the extended-power cascade analysis method, which designs a
hybrid power system with hydrogen storage. The component costs of the
HSS are summarized in Table F1 under the assumptions that the power
losses of conversion and transfer are negligible and the efficiencies of the
storage system components are 100% [4].
The capacity of the renewable energy technologies in each scenario
is determined under the following five assumptions:
(1) The lacking electricity that should be supplied by the renewable
energy is calculated as the forecasted total electricity demand
minus the forecasted electricity generation by renewable energy
technologies (wind power and photovoltaic power).
(2) The necessary capacity of renewable energy facilities is 120% of
the forecasted demand, where this percentage considers the
power margin recommended by the Ministry of Trade, Industry,
and Energy in Korea.
(3) As previously mentioned, the energy supplies of wind power and
photovoltaic power fluctuate based on the weather and seasonal
conditions. Thus, in these scenarios, the electricity generation by
wind farms and photovoltaic power stations is estimated by
considering the variations of electricity over four years.
(4) The lifespan of the energy facilities are 30 years for IGCC, 20
years for fuel cells, 25 years for onshore and offshore wind farms,
and 15 years for photovoltaic power stations [57–59].
(5) The economic costs suggested by EIA [60] and environmental
costs suggested by Refs. [61–63] of each renewable energy
technology are used. EIA [60] suggested economical costs
considering capital costs and operating and management (O&M)
costs according to the nominal capacity of the power plants. In
addition, pollutant emissions from each power plant are
4. Results and discussion
4.1. Forecasting of electricity demand and electricity generation by
renewable energy
Fig. 3 shows representative results of data decomposition of wind
electricity generation using the EMD. The numbers of IMFs are 7, 10,
and 8 for total electricity demand, wind electricity generation, and
photovoltaic electricity generation, respectively. The number of IMFs is
the lowest when the total electricity demand data are used; on the other
hand, this number is the highest when of the decomposed wind elec­
tricity generation data are used. In addition, IMF signals appear with
higher frequency in the cases of wind electricity generation compared to
those of electricity demand and photovoltaic electricity generation. The
higher frequencies of IMFs mainly reflect the random noise information
of the original data, and the number of IMFs is increased according to the
6
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Fig. 3. Original wind electricity generation data, and ten IMFs and one residue, which are computed by EMD-based decomposition of the wind electricity gener­
ation data.
amount of noise. This result suggests that wind energy has the highest
noise intensity among the observed data sets. The reason for such erratic
wind power is the climate and geographical characteristics of Jeju Island
and Korea. Wind speed, which is directly correlated with the amount of
wind electricity generation, is relatively low in the summer compared to
the spring, fall, and winter in Korea, while wind speed in the winter is
the highest [64]. The reasons for wind speed variation include the dif­
ferences in temperature between the land and sea and the wind shear
exponent. The wind speeds from the Pacific in the summer are much
lower than those from Siberia in the winter, because the temperature is
much more different between the land and the sea during the winter. In
addition, the wind shear exponent is relatively high in the summer [65].
Therefore, a robust forecasting model should be implemented to predict
the large fluctuations and erratic energy trends according to time,
weather, and season. In this context, this study utilizes a deep
learning-based forecasting model to capture the characteristics of elec­
tricity demand and generation, as well as predict the electricity gener­
ation with robust performance.
Table 2 shows the electricity demand and renewable energy gener­
ation forecasting results of the statistical models and the deep learning
models based on MASE and MAE as the forecasting models evaluators.
GRU shows the best forecasting performance of the electricity demand,
wind electricity generation, and photovoltaic electricity generation
when evaluated by the MASE. On the other hand, the statistical models
show better forecasting performance than the deep learning models with
respect to the MAE. This result is explained by the characteristics of the
7
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Table 2
Forecasting results of the electricity demand and renewable energy generation; the best forecasting model is indicated with bold font.
Evaluator
MASE
Electricity demand
Wind electricity generation
Photovoltaic electricity generation
MAE
Electricity demand
Wind electricity generation
Photovoltaic electricity generation
Training
Test
Training
Test
Training
Test
Training
Test
Training
Test
Training
Test
model evaluators. Although the MAE is generally used to measure the
ability of a forecasting model, it is a suitable model performance eval­
uator when forecasted and measured data have a uniform distribution
[50]. Wind power and photovoltaic power electricity generation have
time-series attributes that continually increase; this is because Jeju Is­
land actively constructs wind farms and photovoltaic power stations to
generate 100% of their energy from renewable sources. Furthermore, as
shown in the results of the data decomposition step, weather and sea­
sonal conditions disturb the efficiency of renewable power generation
and result in highly erratic data.
Fig. 4 (a)-(c) verify the appropriateness of a low MASE value for GRU
regarding the forecasted wind electricity generation, since GRU fore­
casts the wind electricity generation while reflecting data variation.
However, SARIAM and MLR only follow trend found in the data set.
MASE is applicable for evaluating time-series data, and GRU shows the
lowest MASE value among the conventional statistical models and deep
learning models. Additionally, 95% confidence intervals for each fore­
casting model are depicted as the red area in Fig. 4 (a)–(c). The confi­
dence intervals of the GRU-based forecasting model have a similar trend
of increasing and decreasing with the observed wind electricity gener­
ation, thus, GRU has a relatively lower number of outliers compared to
the other forecasting models. Here, outliers represent the observations
that are not included within the 95% confidence intervals, as depicted
by the forecasting models [66]. On the other hand, the confidence in­
tervals by SARIMA and MLR have relatively small widths since they
roughly follow the trend of wind electricity generation. SARIMA and
MLR have limitations in accurately reflecting the highly varying elec­
tricity data. Therefore, it is inferred that the GRU outperforms the
forecast time-series data. Fig. 5 (a)-(c) show forecasted electricity values
using GRU for the electricity demand, wind, and photovoltaic electricity
generation during 2017. The varied time-series trends of the forecasted
electricity values for demand and generation using the GRU are similar
to those for the observed data set. Furthermore, the observed electricity
demand and the generated renewable energy are mostly located within
the 95% confidence intervals of the GRU-based forecasting models. It is
essential to detect the fluctuating trends of electricity generation via
renewable energy technologies to provide insight into the energy char­
acteristics and design of a realistic energy policy. We select GRU as the
optimal electricity forecasting model because it reflects the time-series
and time-varied characteristics of electricity demand and generation.
Thus, the GRU-based forecasted electricity demand, wind electricity
generation, and photovoltaic electricity generation of 2017 are used to
suggest and analyze the renewable energy scenario in Jeju Island.
GRU
LSTM
DNN
SARIMA
MLR
1.12
1.11
0.19
1.76
0.47
1.80
253.80
337.96
101.01
910.31
21.845
159.52
1.35
2.03
0.50
1.85
0.38
2.03
269.92
447.86
210.97
1022.1
18.482
166.62
1.63
1.59
1.10
2.72
0.55
1.93
398.59
443.42
321.11
915.76
24.911
162.46
1.23
1.23
15.6
6.48
7.54
7.75
449.91
535.48
583.37
871.27
59.408
135.78
2.05
1.90
8.01
7.17
8.46
8.67
397.41
437.53
539.57
795.83
54.765
127.08
wind farm, a photovoltaic power station, and a fuel cell plant. As shown
in Fig. 6 (a), the number of renewable energy facilities should supply at
least 120% of the forecasted electricity demand. In total, 0 onshore wind
farms, 2 offshore wind farms, 5 photovoltaic power stations, and 68 fuel
cell plants are used in scenario 1. The characteristics of the highly erratic
and fluctuating electricity generation are observed due to wind elec­
tricity generation. Accounting for these fluctuations, 52% of the elec­
tricity in total is generated from wind power. Relatively low electricity
generation is observed for the summer season, from 152 days to 243
days in Fig. 6 (a), due to low wind. However, energy demand during the
summer is increased due to high temperatures and the use of air con­
ditioning. Therefore, the incorporation of wind farms should be care­
fully considered. The high capacity of renewable energy facilities is
estimated in scenario 1 to satisfy the electricity demand of Jeju Island.
Highly fluctuating wind electricity generation induces high capacities
for the facilities by overestimating the number of other facilities. Thus,
surplus electricity is generated during other seasons. These results
indicate the necessity of a policy design with careful consideration of
renewable energy and a thorough understanding of the renewable en­
ergy facility characteristics. In this context, this study intends to inform
sustainable energy policies and guide policy makers by suggesting and
analyzing feasible and reliable renewable energy scenarios while
implementing other renewable energy facilities.
Fig. 6 (b) shows the results of scenario 2, which uses additional IGCC
sustainable energy facilities. Scenario 2 generates electricity from 1
IGCC, 1 onshore wind farm, 1 offshore wind farm, 3 photovoltaic power
stations, and 21 fuel cell plants. Compared to the base scenario in Fig. 6
(a), the IGCC generates stable electricity and minimizes the surplus
electricity by simultaneously decreasing unnecessary renewable energy
facilities. Stable electricity generation by the IGCC decreases the total
capacity of renewable energy facilities from 1550 MW to 1290 MW
compared to the current policy. The IGCC generates 40% of the elec­
tricity, and wind farms, which are the main renewable energy sources of
the base scenario, generate 38%. This electricity generation ratio means
that the dependency on wind energy is decreased from 52% to 38%, and
electricity generation is more stable as the seasons change. By dimin­
ishing the capacity of total facilities and wind farms, Fig. 6 (b) shows
that the intensity of surplus electricity is dramatically decreased, and
scenario 2 is expected to reduce operation costs. Therefore, utilization of
an additional sustainable energy facility can be an effective policy so­
lution to achieve a 100% sustainable energy system with economic
benefits.
Scenario 3, which uses HSS to store surplus electricity and then
supply that electricity when needed, is shown in Fig. 6 (c). The number
of estimated facilities is 2 onshore wind farms, 1 offshore wind farm, 3
photovoltaic power stations, and 63 fuel cell plants. A notable result is
that the 85 MW HSS is used in scenario 3. The HSS mainly furnishes
saved electricity for summer from day 152 to day 243. This helps pre­
vent overestimation of the renewable energy capacity, as well as
generate and provide stable electricity during the summer. Thus, the
total capacity, including the facilities and electricity storage in scenario
4.2. Economic-environmental analysis of feasible renewable energy
scenarios
Fig. 6 (a)-(d) show the electricity generation through renewable
energy facilities according to the suggested renewable energy scenario
in Jeju Island during 2017. Scenario 1, which is based on the current
policy of Jeju Island, is composed of an onshore wind farm, an offshore
8
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Fig. 4. Forecasted wind electricity generation and comparison of the observed
data and confidence intervals by using (a) GRU, (b) SARIMA, and (c) MLR; the
days from 1461 days to 1825 days represent the year 2017.
Fig. 5. Results of GRU-based forecasting model and confidence intervals for (a)
electricity demand, (b) wind electricity generation, and (c) photovoltaic elec­
tricity generation; the days from 1461 days to 1825 days represent the
year 2017.
3, is diminished from 1550 MW to 1375 MW compared to scenario 1.
The HSS has a low capacity despite its benefit of preventing capacity
overestimation of renewable energy facilities and providing stable
electricity. The reason for the low HSS capacity is the high price com­
ponents of the system. The polymer electrolyte membrane fuel cell, also
called a proton exchange membrane fuel cell, is utilized in the HSS. The
polymer electrolyte membrane fuel cell is a widely commercialized type
of fuel cell due to its quiet operation, diverse flexible construction and
application, and effective, cleaner chemical-electrical energy
conversion. Nonetheless, it is expensive and has short durability. The
dependency of a platinum catalyst, frail fabrication of the membrane,
and bipolar plate materials are the main obstacles of fuel cell utilization
in renewable energy systems [55]. Although the components of the HSS
are expensive, the storage system’s small capacity reduces the amount of
unnecessary electricity generation. If the economic problems of the HSS
can be overcome, HSS will be useful in Jeju Island’s sustainable energy
9
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Fig. 6. Results of the renewable energy scenarios, including the detailed capacity and ratio of each renewable energy facility, and variations of electricity generation
for 2017; (a) scenario 1, (b) scenario 2, (c) scenario 3, and (d) scenario 4.
10
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
plans and policy.
Both the IGCC and the HSS are utilized in scenario 4, which is shown
in Fig. 6 (d). There is 1 IGCC, 1 onshore wind farm, 1 offshore wind farm,
3 photovoltaic power stations, 21 fuel cell plants, and no use of the HSS.
Contrary to our expectation, the HSS is not implemented in scenario 4
due to its inherent high costs and sufficient stable electricity generation
by the IGCC. The renewable energy facilities of scenario 4 correspond to
those of scenario 2.
Table 3 shows the capacity of the facilities and the economicenvironmental costs of each scenario. Scenario 1, estimated using the
current policy of Jeju Island, shows the highest capacity and costs. These
high values result mainly from wind electricity, which interrupts
effective electricity generation. Therefore, scenarios 2, 3, and 4 are
suggested to inform effective electricity generation systems and the
related policies. Scenario 2 used the IGCC as the main electricity gen­
eration facility. The IGCC uses coal, but scenario 2 still maintains low
environmental costs for operation and maintains consistent renewable
energy. The reason for these low environmental costs is the combination
of mitigation technology CCS and IGCC; it uses a water gas shift reactor
system and a two-stage acid gas removal system to capture CO2 from the
syngas before combustion [60]. This IGCC-CCS effectively prevents
pollutant emission and reduces environmental costs. Moreover, the
IGCC decreases the required electricity capacity of the facilities and
decreases the costs. Compared to the base scenario, scenario 2 decreases
the required capacity of the facilities by 16.74%, economic costs by
33.94%,
environmental
costs
by
52.91%,
and
total
economic-environmental costs by 36.16%. Thus, scenario 2 saves $326,
490,000 dollars per year while satisfying sufficient electricity generation
for electricity consumption in Jeju Island. The effect of scenario 2 is also
highlighted in Fig. 7 considering the capacity and cost aspects. Scenario
3, which uses an HSS to save and provide surplus electricity, decreases
the required capacity of the facilities by 11.29% and the total
economic-environmental costs by 8.26%. Though scenario 3 reduces the
economic-environmental costs compared to the base scenario, the costs
are higher than those of scenario 2.
To summarize the results of the four scenarios, scenario 2 is the best
scenario to guide policy and attain a renewable energy generation sys­
tem on Jeju Island. In addition, IGCC is an effective renewable energy
facility selection with respect to economic and environmental aspects.
We additionally assess the optimal scenario by conducting a technoeconomic-environmental SWOT analysis to evaluate the feasibility and
reliability of applying it to Jeju Island’s energy system according to the
renewable energy conditions of both Korea and the world.
4.3. SWOT analysis considering techno-economic-environmental aspects
for the optimal renewable energy scenario
The strengths represent the available resources for enhancing the
performance of the scenario. The main strength of the optimal scenario
is its application of IGCC. IGCC is a potential clean coal power genera­
tion process that has two main characteristics—sustainability and flex­
ibility. First, coal is mainly used in the IGCC. Globally, the reserve of coal
is 59.9%, oil is 23.4%, and natural gas is 16.7%. In addition, the
reserves-to-production ratio, or the amount of remaining fuel expressed
over time, is the longest for coal (109), followed by natural gas (55.7)
and oil (52.9) [67]. Therefore, the abundant reserve of coal supports the
suitability of IGCC. Second, IGCC is operated under flexible multiple
feeds and generates multiple products. The IGCC uses biomass from
agriculture, forestry, and energy crops; petroleum residues, which
include coke, asphalt, and heavy oil; as well as coal such as peat, lignite,
and anthracite as fuel [56]. The diverse feeds lend flexibility to the IGCC.
These two technical characteristics enhance the benefits attained by the
IGCC. Further, the stable and steady electricity generation of the IGCC is
better than the erratic and fluctuating electricity generation that results
from wind and photovoltaic power. Thus, the optimal scenario has many
advantages with regard to raw materials, stable electricity, and low total
economic-environmental costs.
However, IGCC has weaknesses that can diminish its competitive
advantages and efficiency. First, the IGCC emits pollutants. Although the
IGCC is a clean power process and is combined with the CCS to prevent
CO2 emission, the IGCC is a fossil fuel-based power plant. Thus, the IGCC
with CCS emits SO2, NOX, and CO2, which violates the zero carbon
emission objective of Jeju Island’s energy policy. Second, this scenario
does not utilize the abundant wind source of Jeju Island. In its 1290 MW
capacity, 38% of the electricity is supplied by onshore and offshore wind
farms. Jeju Island has been evaluated as the best location in Korea for
operating wind farms due to its abundant and strong wind speeds [68].
Despite the geographical advantages of wind, non-stable electricity
generation from wind farms prevents the effective utilization of wind
power on Jeju Island. New technologies to increase the efficiency of
wind turbine generators may overcome this limitation of wind power.
External policies are additional contributing opportunities for sus­
tainable energy in Korea. First, the Korean ministry announced an IGCC
construction schedule considering long-term electricity supply and de­
mand. Based on this construction plan, a 300 MW IGCC facility was
constructed at 2016 in Taean, Korea. Moreover, an additional plan was
established in 2015 to build an IGCC plant in Namhae, with expected
completion in 2022. Thus, a positive prospective of the IGCC is expected
in Korea. Second, the Moon administration emphasizes phasing out
nuclear power and closing old coal power plants to assure public safety
and promote environmentally friendly practices. According to the newly
announced policy, the 40-year Kori-1 nuclear power reactor was retired,
and five 30-year old coal power plants will be closed. Therefore, alter­
native feasible clean power plants are needed to compensate for the
closure of conventional power plants and supply stable electricity gen­
eration. This may be an opportunity to develop an IGCC plant as well as
other sustainable energy technologies.
Threats, which are outside factors, affect the energy supply and de­
mand and cause problems. Coal, which is the basic fuel used to operate
the IGCC plant, is mostly imported. Only 1.3% of coal is mined in Korea,
and 98.7%, which is 113.5 million tons of coal, is imported from Austria,
Indonesia, Russia, and Canada [44]. Although coal is mined in several
regions throughout the world and has the advantage of energy security
compared to oil, which is provided from limited regions such as the
Middle East, the high dependency on coal imports may threaten the
dependability of the IGCC operation. Another threat to the IGCC on Jeju
Island is the shale gas industry. Shale gas production is increasing and
will soon constitute half of the total amount of natural gas; thus, the
price of shale gas is continuously decreasing. Furthermore, the Korean
government plans to utilize shale gas by promoting manpower and
Table 3
Comparison of the detailed capacity and total annual costs for each scenario.
Onshore wind
farms (MW)
Offshore wind
farms (MW)
Photovoltaic
power stations
(MW)
Fuel cell plants
(MW)
IGCC (MW)
HSS (MW)
Total capacity
(MW)
Economic costs
($/year)
Environmental
costs ($/year)
Total costs
($/year)
Scenario 1
Scenario 2
Scenario 3
Scenario 4
–
100
200
100
800
400
400
400
100
60
60
60
650
210
630
210
–
–
1550
520
–
1290
–
85
1375
520
–
1290
797,461,120
526,797,936
726,155,524
526,797,936
105,510,000
49,683,000
102,260,000
49,683,000
902,970,000
576,480,000
828,420,000
576,480,000
11
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Fig. 7. Effect of the optimal renewable energy scenario with respect to the diminished total capacity of electricity generation facilities and the decreased economicenvironmental costs.
Research & Development [69]. Interest in shale gas may restrain the
expansion of IGCC and renewable energy technologies in Korea. The
SWOT analysis results of the optimal scenario for Jeju Island with a
100% renewable energy system are summarized in Table 4.
forecasting model showed the best performance in terms of the MASE
when compared to other deep learning models and statistical models.
Moreover, compared to the other models, the GRU has the ability to
reflect and forecast data that rapidly fluctuate over time. The suggested
forecasting model provides feasible grounds for establishing appropriate
energy policies. Black out, which occurs when electricity generation is
insufficient, and surplus electricity, which is when electricity generation
exceeds the capacity of the system, can be prevented via robust policies
based on an accurate energy forecasting model.
Based on the forecasted electricity demand and generation by GRU,
four feasible renewable energy scenarios consisting of wind farms,
photovoltaic power stations, fuel cell plants, IGCC, and/or HSS are
suggested and evaluated according to various economic and environ­
mental aspects. Scenario 2, which uses the IGCC, wind farms, photo­
voltaic power stations, and fuel cell plants, was determined as the best
scenario. The results illustrate that scenario 2 generates stable electricity
by consistently providing 520 MW electricity from a IGCC with CCS
system, 100 MW electricity from onshore wind farms, 400 MW elec­
tricity from offshore wind farms, 60 MW electricity from photovoltaic
power stations, and 210 MW electricity from fuel cell plants. Scenario 2
also reduces the total economic-environmental costs by 36.16%
compared to the existing policy of Jeju Island.
The SWOT assessment explains techno-economic-environmental
circumstances to expand the suggested renewable energy technologies.
Additionally, it can be extended to analyze scenarios in light of different
system aspects such as robust carbon emission reduction. Therefore, this
approach can assess and establish other renewable energy-related sys­
tems and policy implementations. Consequently, through sequential
forecasting, renewable energy scenarios, and SWOT assessment, this
study provides guidelines for the design of renewable energy systems
and policies by considering time-varying energy trends, specific
5. Conclusions
We suggested and analyzed various renewable energy scenarios to
adhere to the objectives of Jeju Island to attain 100% sustainable
environmental electricity generation and inform policy. We developed a
deep learning-based forecasting model, which is seldom used for energy
forecasting, to overcome the performance limitations of conventional
forecasting models and enhance the model performance. The GRU
Table 4
The summarized SWOT results of the optimal renewable energy scenario of Jeju
Island.
Strengths
Weaknesses
- Sustainability and flexibility of the IGCC
(technical)
- Stable and steady electricity generation
not affected by seasonal and weather
conditions (technical)
- Reduction of total economicenvironmental costs (economic and
environmental)
Opportunities
- Difficult to maintain zero carbon
emission (technical and
environmental)
- Lacking utilization of abundant
wind sources of Jeju Island
(technical)
- Vitalization of IGCC use in Korea
(technical)
- Alternative and effective energy source
to nuclear power plants and old coal
power plants (technical and economic)
- High dependency on fuel import
(economic)
- Rise of shale gas potential in the
world (economic)
Threats
12
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
characteristics of renewable energy technologies, and diagnosis of
techno-economic-environmental domestic and global energy trends.
Writing - review & editing. ChangKyoo Yoo: Conceptualization, Vali­
dation, Supervision.
CRediT authorship contribution statement
Acknowledgements
KiJeon Nam: Conceptualization, Methodology, Software, Formal
analysis, Investigation, Writing - original draft, Writing - review &
editing, Visualization. Soonho Hwangbo: Conceptualization, Method­
ology, Software, Formal analysis, Resources, Writing - original draft,
This work was supported by a National Research Foundation of
Korea (NRF) grant funded by the Korean government (MSIP) (no.
2017R1E1A1A03070713).
Appendix A
A. Current energy system status and energy policy of Jeju island
Jeju Island, located in the south sea of Korea, announced a sustainable energy policy that replaces conventional fossil fuel power plants with
sustainable energy technologies such as wind power, photovoltaic power, and fuel cells. The current energy system of Jeju Island generated only
22.29% of its electricity through onshore wind farms and photovoltaic power stations. The ultimate objective of the current policy is to achieve 100%
sustainable energy [1]. The proportions of electricity generation facilities and the location of current power systems are respectively shown in Figs A1
(a) and (b). In total, 42.17% of electricity is supplied by high voltage direct current (HVDC), 35.54% of electricity is generated by thermal power
plants, and 22.29% of electricity is generated by renewable energy technologies. Among the renewable energy sources constituting 22.29% of total
electricity generation, 80.90% and 19.10% of the electricity is generated by onshore wind farms and photovoltaic power stations, respectively [1]. The
first objective of this sustainable energy policy is the installation of wind farms to utilize the wind energy sources of Jeju Island. Unfortunately,
predominately focusing on wind farms has resulted in an inescapable dependency on fossil fuel energy sources. Therefore, a 700 MW HVDC was
installed and has been used to supply electricity to Jeju Island; the facilities are located north of the island and receive their electricity from inland
Korea. The main electricity generation facilities on Jeju Island are thermal power plants, which supply 61.45% of the electricity (not considering
transmitted electricity). These percentages show that the current electricity generation is still highly dependent on thermoelectric power plants.
Therefore, careful and detailed consideration of new renewable energy centers is needed to achieve 100% sustainable energy.
13
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
Fig. A.1. Energy generation system of Jeju Island. (a) Ratio of electricity generation and (b) location of electricity generation facilities.
The objective of this policy is to provide a renewable energy system composed of 2 GW offshore wind farms, 300 MW onshore wind farms, 100 MW
photovoltaic power stations, and 520 MW fuel cell plants. Wind power is the main source of energy in this policy, and has fluctuating electricity
generation characteristics that interrupt sustainable and stable electricity supply. Therefore, the energy policy should carefully consider the char­
acteristics of Jeju Island’s climate as well as efficient electricity generation technologies. To achieve 100% renewable energy technologies for elec­
tricity, we suggest an energy forecasting model that considers a time-varying electricity trend and feasible renewable energy scenarios to design an
acceptable sustainable energy policy for Jeju Island.
B. Empirical mode decomposition (EMD)
The main algorithm of EMD is as follows [2–4].
14
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
(1) Identify the local maxima and minima of the data xðtÞ. Then, connect all the local maxima with a cubic spline curve to produce the upper
envelope. Calculate the difference function hðtÞ between xðtÞ and the mean value of the upper and lower envelopes mðtÞ.
(2) hk ðtÞ, which is the difference function at iteration k, shows a zero mean, so hk ðtÞ is then designated as the first IMF1 ðtÞ. Obtain the residue r1 ðtÞ
by subtracting IMF1 ðtÞ from xðtÞ. Consider r1 ðtÞ as the new original data. Next, calculate IMF2 ðtÞ, since r1 ðtÞ includes information from a longer
period. Iterate the calculation of rn ðtÞ until rn ðtÞ becomes a monotonic function from which IMFðtÞ can no longer be extracted.
(3) Finally, the original data xðtÞ is decomposed into the residue rn ðtÞ and the sum of IMF1 ðtÞ to IMFn ðtÞ as in Eq. (B.1).
n
X
xðtÞ ¼
(B.1)
IMFi ðtÞ þ rn ðtÞ
i¼1
C. Conventional statistical forecasting models
The general form of the MLR is Eq. (C.1).
(C.1)
y ¼ β 0 þ β 1 x1 þ β 2 x2 þ ⋯ þ β n xn
here, βi is the corresponding regression coefficient estimated by the least squares method.
The SARIMA is generally represented by SARIMA (p,d,q)(P,D,Q)(s) where p is the autoregressive order, d is the differencing level, q is the moving
average order, P is the seasonal autoregressive notation, D is the seasonal integrated notation, Q is the moving average notation, and s is the length of
seasonal period. The Box-Jenkins methodology is utilized in the SARIMA model development and consists of four steps: identification, estimation,
diagnostic checking, and forecasting. A detailed model description with mathematical equations can be found in Jeong et al. [5] and Vagropoulos et al.
[6].
D. Deep learning-based forecasting models
LSTM has memory block units located inside recurrent hidden layers. The units have memory cells that save temporal information and gates that
input, output, and forget information as it flows through the system. The input gate controls input data into the memory cell, the output gate controls
output data into the next network, and the forget gate controls the internal state of the memory cell by preventing continuous input streams through
adaptively forgetting or resetting the memory of the cell [7,8]. The equations of the LSTM transition are expressed in Eqs. (D.1)-(D.6):
�
it ¼ σ W ðiÞ xt þ U ðiÞ ht 1 þ bðiÞ
(D.1)
ft ¼ σ W ðf Þ xt þ U ðf Þ ht
ot ¼ σ ðW ðoÞ xt þ U ðoÞ ht
1
�
(D.2)
þ bðoÞ Þ
(D.3)
þ bðf Þ
1
ut ¼ tanhðW ðuÞ xt þ U ðuÞ ht
ct ¼ it ∘ut þ ft ∘ct
1
(D.4)
þ bðuÞ Þ
(D.5)
1
(D.6)
ht ¼ ot ∘tanhðct Þ
where it , ft , and ot are the input, forget, and output gates, respectively, ct is the memory cell, ht is the output of hidden state, xt is the input at the
current time, W, U and b are the parameter matrices and vector, σ is the logistic sigmoid function, and ∘ is elementwise multiplication.
GRU, proposed by Cho et al. [9], is an alternative technique to LSTM. The main difference between GRU and LSTM is that a single gating unit in
GRU simultaneously controls the forgetting factor and the decision to update the state unit; a separate memory cell therefore does not exist in GRU
[10]. GRU is expressed by the following Eqs. (D.7)-(D.10)
rt ¼ σðW ðrÞ xt þ U ðrÞ ht
1
þ bðrÞ Þ
(D.7)
zt ¼ σðW ðzÞ xt þ U ðzÞ ht
1
þ bðzÞ Þ
(D.8)
h~t ¼ tanh W ðhÞ xt þ U ðhÞ ðrt ∘ ht 1 Þ þ bðhÞ
ht ¼ zt ∘ht
1
þ ð1
�
(D.9)
(D.10)
zt Þ∘~
ht
~t is the candidate output of hidden state. To balance the previous output of hidden state ht and the
where rt is the reset gate, zt is the update gate, and h
~t , the input and forget gates are combined into the update gate zt . Then, the reset gate rt controls the previous state
candidate output of hidden state h
[11].
E. Descriptions of renewable energy technologies
The wind farms receive wind power through the rotor and gearbox, and convert mechanical power into electrical power using a power converter
and transformer. Although wind power is clean energy with no pollutant emission, electricity generated from wind turbines is highly erratic and
fluctuates, making it difficult to implement wind farms into energy system plans [12]. Photovoltaic power stations consist of many cells connected in
15
Renewable and Sustainable Energy Reviews 122 (2020) 109725
K. Nam et al.
series and parallel that convert solar power into electrical power. Similar to the wind farms, the photovoltaic power stations can reduce energy
consumption of fossil fuels and mitigate air pollution. However, photovoltaic power stations are dependent on the weather [13,14]. The fuel cell
plants, composed of a fuel electrode, an oxidant electrode, and an electrolyte, produce energy from hydrogen oxidization. These are regarded as
flexible chemical-to-electrical energy converters with efficient and clean advantages. Despite the advantages, the drawback of fuel cells is the high cost
of equipment including catalysts, membranes, and bipolar plates [15].
The IGCC, evaluated as a clean coal power plant, generates energy from coal using a combined cycle. The combined cycle consists of a gas turbine
and a steam turbine. The gas turbine combusts syngas by combing it with coal, a water slurry, and oxygen under high temperatures, and the steam
turbine produces electricity using the high pressure stream generated by excess heat from the syngas combustion. Compared to a conventional
pulverized coal combustion process, the advantages of IGCC are that it is associated with 10–30% pollutant emission at equivalent power generation,
30% less water use, lower fused slag production, higher efficiency, and higher energy output. In addition, carbon capture and sequestration (CCS) can
be adopted into IGCC to mitigate CO2 emission [16,17].
The HSS, which is generally used as an energy storage system, stores immoderate generated electricity in a hydrogen storage tank via deionization
and electrolysis. When generated electricity is lower than the electricity demand, the HSS delivers saved energy to a fuel cell that produces energy and
water from the saved hydrogen. The high energy density of the hydrogen is one advantage of the HSS [12,18].
F. Economic and environmental cost
Table F.1
Characteristics of renewable energy facilities including cost information and environmental impacts [18,19].
Facility
Fuel
Nominal
capacity
(kW)
Nominal
heat rate
(Btu/kWh)
Capital cost
($">$/kW)
O&">&M
($">$/kW)
SO2 (lb/
MMBtu)
NOX (lb/
MMBtu)
CO2 (lb/
MMBtu)
Description
60 wind turbine generators, each with
a rated capacity of 1.5 MW
80 wind turbine generators, each with
a rated capacity of 5 MW
40 half-megawatt building blocks,
each block consisting of groups of PV
modules connected to a 500 KW AC
inverter
Multiple phosphoric acid fuel cell
units, each with a power output of
400 kW, for a total output of 10 MW
Two combustion turbine (60 Hz
machines rated at 255MVA with an
18 kV output) and one steam turbine
(60 Hz machines rated at 333MVA
with 18 kV output)
Components of the hydrogen storage
system to store and provide electricity
Onshore wind
farm
Offshore wind
farm
Photovoltaic
power station
Wind
100,000
–
2328
41.60
0
0
0
Wind
400,000
–
6554
77.84
0
0
0
Solar
20,000
–
4400
29.19
0
0
0
Fuel cell plant
Gas
10,000
9500
7477
396.21
0.00013
0.013
130
IGCC with CCS
Coal
520,000
10,700
6942
154.48
0.015
0.0075
20.6
Polymer
electrolyte
membrane fuel
cell
Electrolyzer
Hydrogen tank
Hydrogen
–
–
2500
62.5
–
–
–
Hydrogen
Hydrogen
–
–
–
–
100
30
2
0.15
–
–
–
–
–
–
Table F.2
Environmental cost of the pollutant emission [20–22].
Pollutant
Environmental cost ($">$/ton/yr)
Considered environmental damaged effect
SO2
NOX
CO2
1494.71
296.66
34.80
Premature mortalities, illness, reduced timber and crops yields, reduced visibility effect, and depreciation of man-made materials
Agriculture, energy production, water availability, human health, coastal communities, and biodiversity
References
[6] Choi S, Oh J, Hwang Y, Lee H. Life cycle climate performance evaluation (LCCP) on
cooling and heating systems in South Korea. Appl Therm Eng 2017;120:88–98.
[7] Kim S, Lee H, Kim H, Jang DH, Kim HJ, Hur J, et al. Improvement in policy and
proactive interconnection procedure for renewable energy expansion in South
Korea. Renew Sustain Energy Rev 2018;98:150–62.
[8] Kim H, Baek S, Park E, Chang HJ. Optimal green energy management in Jeju, South
Korea - on-grid and off-grid electrification. Renew Energy 2014;69:123–33.
[9] Kwon S, Won W, Kim J. A superstructure model of an isolated power supply system
using renewable energy: development and application to Jeju Island, Korea. Renew
Energy 2016;97:177–88.
[10] Park E. Potentiality of renewable resources: economic feasibility perspectives in
South Korea. Renew Sustain Energy Rev 2017;79:61–70.
[11] Lee J, Park GL, Kim EH, Kim YC, Lee IW. Wind speed modeling based on artificial
neural networks for Jeju area. Int J Control Autom 2012;5:81–8.
[12] Wang Y, Wang J, Zhao G, Dong Y. Application of residual modification approach in
seasonal ARIMA for electricity demand forecasting: a case study of China. Energy
Pol 2012;48:284–94.
[1] Foster E, Contestabile M, Blazquez J, Manzano B, Workman M, Shah N. The
unstudied barriers to widespread renewable energy deployment: fossil fuel price
responses. Energy Pol 2017;103:258–64.
[2] Pachauri RK, Allen MR, Barros VR, Broome J, Cramer W, Christ R, et al. Climate
change 2014: synthesis report. Contribution of working groups I, II and III to the
fifth assessment report of the intergovernmental panel on climate change. IPCC;
2014.
[3] Chu S, Cui Y, Liu N. The path towards sustainable energy. Nat Mater 2016;16:
16–22.
[4] Janghorban Esfahani I, Lee SC, Yoo CK. Extended-power pinch analysis (EPoPA)
for integration of renewable energy systems with battery/hydrogen storages.
Renew Energy 2015;80:1–14.
[5] Hwangbo S, Nam KJ, Han J, Lee IB, Yoo CK. Integrated hydrogen supply networks
for waste biogas upgrading and hybrid carbon-hydrogen pinch analysis under
hydrogen demand uncertainty. Appl Therm Eng 2018;140:386–97.
16
Renewable and Sustainable Energy Reviews 122 (2020) 109725
K. Nam et al.
[46] Connolly D, Lund H, Mathiesen BV. Smart Energy Europe: the technical and
economic impact of one potential 100% renewable energy scenario for the
European Union. Renew Sustain Energy Rev 2016;60:1634–53.
[47] Bale CSE, Varga L, Foxon TJ. Energy and complexity: new ways forward. Appl
Energy 2015;138:150–9.
[48] Fan C, Xiao F, Wang S. Development of prediction models for next-day building
energy consumption and peak power demand using data mining techniques. Appl
Energy 2014;127:1–10.
[49] Awad MA, Khan LR. Web navigation prediction using multiple evidence
combination and domain knowledge. IEEE Trans Syst Man, Cybern Part ASystems
Humans 2007;37:1054–62.
[50] Chai T, Draxler RR. Root mean square error (RMSE) or mean absolute error (MAE)?
-Arguments against avoiding RMSE in the literature. Geosci Model Dev (GMD)
2014;7:1247–50.
[51] Hyndman RJ, Koehler AB. Another look at measures of forecast accuracy. Int J
Forecast 2006;22:679–88.
[52] Asadi S, Amiri SS, Mottahedi M. On the development of multi-linear regression
analysis to assess energy consumption in the early stages of building design. Energy
Build 2014;85:246–55.
[53] Lv Y, Yu M, Jiang G, Shao F, Peng Z, Chen F. No-reference stereoscopic image
quality assessment using binocular self-similarity and deep neural network. Signal
Process Image Commun 2016;47:346–57.
[54] G�
eron A. Hands-on machine learning with Scikit-Learn and TensorFlow: concepts,
tools, and techniques to build intelligent systems. California: O’Reilly Media; 2019.
[55] Sharaf OZ, Orhan MF. An overview of fuel cell technology: fundamentals and
applications. Renew Sustain Energy Rev 2014;32:810–53.
[56] Parraga J, Khalilpour KR, Vassallo A. Polygeneration with biomass-integrated
gasification combined cycle process: review and prospective. Renew Sustain
Energy Rev 2018;92:219–34.
[57] Zhaofeng X, Hetland J, Kvamsdal HM, Zheng L, Lianbo L. Economic evaluation of
an IGCC cogeneration power plant with CCS for application in China. Energy
Procedia 2011;4:1933–40.
[58] Wierzbowski M, Lyzwa W, Musial I. MILP model for long-term energy mix planning
with consideration of power system reserves. Appl Energy 2016;169:93–111.
[59] Lyzwa W, Wierzbowski M. Load duration curve in the long-term energy mix
optimization. Int Conf Eur Energy Mark 2016 13th;2016:1–5.
[60] Eia US. Updated capital cost estimates for utility scale electricity generating plants.
US Energy Inf Adm 2013;524.
[61] Huang C-H, Bagdon BA. Quantifying environmental and health benefits of using
woody biomass for electricity generation in the Southwestern United States. J For
Econ 2018;32:123–34.
[62] Marten AL, Newbold SC. Estimating the social cost of non-CO2 GHG emissions:
methane and nitrous oxide. Energy Pol 2012;51:957–72.
[63] Muller NZ, Mendelsohn R. Efficient pollution regulation: getting the prices right.
Am Econ Rev 2009;99:1714–39.
[64] Oh K-Y, Kim J-Y, Lee J-K, Ryu M-S, Lee J-S. An assessment of wind energy potential
at the demonstration offshore wind farm in Korea. Energy 2012;46:555–63.
[65] Ali S, Lee S-M, Jang C-M. Statistical analysis of wind characteristics using Weibull
and Rayleigh distributions in Deokjeok-do Island–Incheon, South Korea. Renew
Energy 2018;123:652–63.
[66] Hwangbo S, Nam KJ, Heo SK, Yoo CK. Hydrogen-based self-sustaining integrated
renewable electricity network (HySIREN) using a supply-demand forecasting
model and deep-learning algorithms. Energy Convers Manag 2019;185:353–67.
[67] Xiang D, Qian Y, Man Y, Yang S. Techno-economic analysis of the coal-to-olefins
process in comparison with the oil-to-olefins process. Appl Energy 2014;113:
639–47.
[68] Kim E-H, Kim J-H, Kim S-H, Choi J, Lee KY, Kim H-C. Impact analysis of wind
farms in the Jeju Island power system. IEEE Syst J 2011;6:134–9.
[69] Chen W-M, Kim H, Yamaguchi H. Renewable energy in eastern Asia: renewable
energy policy review and comparative SWOT analysis for promoting renewable
energy in Japan, South Korea, and Taiwan. Energy Pol 2014;74:319–29.
[13] Chang W-Y. A literature review of wind forecasting methods. J Power Energy Eng
2014;2:161–8.
[14] Suganthi L, Samuel AA. Energy models for demand forecasting - a review. Renew
Sustain Energy Rev 2012;16:1223–40.
[15] Li Q, Loy-Benitez J, Nam K, Hwangbo S, Rashidi J, Yoo C. Sustainable and reliable
design of reverse osmosis desalination with hybrid renewable energy systems
through supply chain forecasting using recurrent neural networks. Energy 2019;
178:277–92.
[16] Loy-Benitez J, Vilela P, Li Q, Yoo C. Sequential prediction of quantitative health
risk assessment for the fine particulate matter in an underground facility using
deep recurrent neural networks. Ecotoxicol Environ Saf 2019;169:316–24.
[17] Thatcher MJ. Modelling changes to electricity demand load duration curves as a
consequence of predicted climate change for Australia. Energy 2007;32:1647–59.
[18] Bianco V, Manca O, Nardini S. Electricity consumption forecasting in Italy using
linear regression models. Energy 2009;34:1413–21.
[19] Goia A, May C, Fusai G. Functional clustering and linear regression for peak load
forecasting. Int J Forecast 2010;26:700–11.
[20] Barak S, Sadegh SS. Forecasting energy consumption using ensemble ARIMAANFIS hybrid algorithm. Int J Electr Power Energy Syst 2016;82:92–104.
[21] Reikard G. Predicting solar radiation at high resolutions: a comparison of time
series forecasts. Sol Energy 2009;83:342–9.
[22] Palomares-Salas JC, De La Rosa JJG, Ramiro JG, Melgar J, Agüera A, Moreno A.
ARIMA vs. neural networks for wind speed forecasting. IEEE Int Conf Comput Intell
Meas Syst Appl 2009;2009:129–33.
[23] Li LL, Li JH, He PJ, Wang CS. The use of wavelet theory and ARMA model in wind
speed prediction. Int Conf Electr Power Equip - Switch Technol 2011 1st;2011:
395–8.
[24] Han Hejase, Assi AH. Time-series regression model for prediction of mean daily
global solar radiation in Al-ain, UAE. ISRN Renew Energy 2012;2012.
[25] Vagropoulos SI, Chouliaras GI, Kardakos EG, Simoglou CK, Bakirtzis AG.
Comparison of SARIMAX, SARIMA, modified SARIMA and ANN-based models for
short-term PV generation forecasting. IEEE Int Energy Conf 2016;2016:1–6.
[26] Wang C, Grozev G, Seo S. Decomposition and statistical analysis for regional
electricity demand forecasting. Energy 2012;41:313–25.
[27] Geem ZW, Roper WE. Energy demand estimation of South Korea using artificial
neural network. Energy Pol 2009;37:4049–54.
[28] Ekonomou L. Greek long-term energy consumption prediction using artificial
neural networks. Energy 2010;35:512–7.
[29] Kaytez F, Taplamacioglu MC, Cam E, Hardalac F. Forecasting electricity
consumption: a comparison of regression analysis, neural networks and least
squares support vector machines. Int J Electr Power Energy Syst 2015;67:431–8.
[30] Muralitharan K, Sakthivel R, Vishnuvarthan R. Neural network based optimization
approach for energy demand prediction in smart grid. Neurocomputing 2018;273:
199–208.
[31] Li G, Shi J. On comparing three artificial neural networks for wind speed
forecasting. Appl Energy 2010;87:2313–20.
[32] Chang WY. Wind energy conversion system power forecasting using radial basis
function neural network. Appl Mech Mater 2013:1067–71.
[33] Troncoso A, Salcedo-Sanz S, Casanova-Mateo C, Riquelme JC, Prieto L. Local
models-based regression trees for very short-term wind speed prediction. Renew
Energy 2015;81:589–98.
� Fern�
[34] Gala Y, A
andez, Díaz J, Dorronsoro JR. Hybrid machine learning forecasting
of solar radiation values. Neurocomputing 2016;176:48–59.
[35] Heinermann J, Kramer O. Machine learning ensembles for wind power prediction.
Renew Energy 2016;89:671–9.
[36] Voyant C, Notton G, Kalogirou S, Nivet ML, Paoli C, Motte F, et al. Machine
learning methods for solar radiation forecasting: a review. Renew Energy 2017;
105:569–82.
[37] Wang H, qiang Li G, Wang G bing, Peng J, Jiang H, Liu Y. Deep learning based
ensemble approach for probabilistic wind power forecasting. Appl Energy 2017;
188:56–70.
[38] Kim J, Park H-I. Policy directions for the smart grid in Korea. IEEE Power Energy
Mag 2010;9:40–9.
[39] Aslani A, Helo P, Naaranoja M. Role of renewable energy policies in energy
dependency in Finland: system dynamics approach. Appl Energy 2014;113:
758–65.
[40] Scarlat N, Dallemand JF, Monforti-Ferrario F, Banja M, Motola V. Renewable
energy policy framework and bioenergy contribution in the European union - an
overview from national renewable energy action plans and progress reports.
Renew Sustain Energy Rev 2015;51:969–85.
[41] He Y, Xu Y, Pang Y, Tian H, Wu R. A regulatory policy to promote renewable
energy consumption in China: review and future evolutionary path. Renew Energy
2016;89:695–705.
[42] Han D, Baek S. Status of renewable capacity for electricity generation and future
prospects in Korea: global trends and domestic strategies. Renew Sustain Energy
Rev 2017;76:1524–33.
[43] Mathiesen BV, Lund H, Connolly D, Wenzel H, Østergaard PA, M€
oller B, et al.
Smart Energy Systems for coherent 100% renewable energy and transport
solutions. Appl Energy 2015;145:139–54.
[44] Cho S, Kim J. Feasibility and impact analysis of a renewable energy source (RES)based energy system in Korea. Energy 2015;85:317–28.
[45] Tripathi L, Mishra AK, Dubey AK, Tripathi CB, Baredar P. Renewable energy: an
overview on its contribution in current energy scenario of India. Renew Sustain
Energy Rev 2016;60:226–33.
Appendix references
[1a] Province JSS-G. Carbon free island Jeju by 2030. Pol Rep 2012.
[2a] Huang NE, Shen Z, Long SR, Wu MC, Snin HH, Zheng Q, et al. The empirical mode
decomposition and the Hubert spectrum for nonlinear and non-stationary time
series analysis. Proc R Soc A Math Phys Eng Sci 1998;454:903–95.
[3a] Alickovic E, Kevric J, Subasi A. Performance evaluation of empirical mode
decomposition, discrete wavelet transform, and wavelet packed decomposition for
automated epileptic seizure detection and prediction. Biomed Signal Process Contr
2018;39:94–102.
[4a] Wang Y, Wu L. On practical challenges of decomposition-based hybrid forecasting
algorithms for wind speed and solar irradiation. Energy 2016;112:208–20.
[5a] Jeong K, Koo C, Hong T. An estimation model for determining the annual energy
cost budget in educational facilities using SARIMA (seasonal autoregressive
integrated moving average) and ANN (artificial neural network). Energy 2014;71:
71–9.
[6a] Vagropoulos SI, Chouliaras GI, Kardakos EG, Simoglou CK, Bakirtzis AG.
Comparison of SARIMAX, SARIMA, modified SARIMA and ANN-based models for
short-term PV generation forecasting. IEEE Int. Energy Conf 2016;2016:1–6.
[7a] Sak H, Senior A, Beaufays F. Long short-term memory recurrent neural network
architectures for large scale acoustic modeling. Proc Annu Conf Int Speech
Commun Assoc 2014.
17
K. Nam et al.
Renewable and Sustainable Energy Reviews 122 (2020) 109725
[16a] Parraga J, Khalilpour KR, Vassallo A. Polygeneration with biomass-integrated
gasification combined cycle process: review and prospective. Renew Sustain
Energy Rev 2018;92:219–34.
[17a] Hossein Sahraei M, McCalden D, Hughes R, Ricardez-Sandoval LA. A survey on
current advanced IGCC power plant technologies, sensors and control systems.
Fuel 2014;137:245–59.
[18a] Janghorban Esfahani I, Lee SC, Yoo CK. Extended-power pinch analysis (EPoPA)
for integration of renewable energy systems with battery/hydrogen storages.
Renew Energy 2015;80:1–14.
[19a] Eia US. Updated capital cost estimates for utility scale electricity generating
plants. US Energy Inf Adm 2013;524.
[20a] Huang C-H, Bagdon BA. Quantifying environmental and health benefits of using
woody biomass for electricity generation in the Southwestern United States. J For
Econ 2018;32:123–34.
[21a] Marten AL, Newbold SC. Estimating the social cost of non-CO2 GHG emissions:
methane and nitrous oxide. Energy Pol 2012;51:957–72.
[22a] Muller NZ, Mendelsohn R. Efficient pollution regulation: getting the prices right.
Am Econ Rev 2009;99:1714–39.
[8a] Tai KS, Socher R, Manning CD. Improved semantic representations from treestructured long short-Term memory networks. 2015. p. 1503. arXiv preprint arXiv.
[9a] Cho K, van Merrienboer B, Bahdanau D, Bengio Y. On the properties of neural
machine translation: encoder–decoder approaches. 2015. p. 1409. arXiv preprint
arXiv.
[10a] Goodfellow I, Bengio Y, Courville A. Deep learning. Cambridge: MIT press; 2016.
[11a] Wu Z, King S. Investigating gated recurrent networks for speech synthesis. Int
Conf Acoust Speech Signal Process 2016:5140–4.
[12a] Díaz-Gonz�
alez F, Sumper A, Gomis-Bellmunt O, Villaf�
afila-Robles R. A review of
energy storage technologies for wind power applications. Renew Sustain Energy
Rev 2012;16:2154–71.
[13a] Onar OC, Uzunoglu M, Alam MS. Modeling, control and simulation of an
autonomous wind turbine/photovoltaic/fuel cell/ultra-capacitor hybrid power
system. J Power Sources 2008;185:1273–83.
[14a] Desideri U, Proietti S, Zepparelli F, Sdringola P, Bini S. Life Cycle Assessment of a
ground-mounted 1778kWp photovoltaic plant and comparison with traditional
energy production systems. Appl Energy 2012;97:930–43.
[15a] Sharaf OZ, Orhan MF. An overview of fuel cell technology: fundamentals and
applications. Renew Sustain Energy Rev 2014;32:810–53.
18
Download