PLOS ONE

RESEARCH ARTICLE

Accurate prediction of electricity consumption

using a hybrid CNN-LSTM model based on
multivariable data
Jaewon Chung1, Beakcheol Jang ID2*

1 Graduate School of International Studies, Yonsei University, Seoul, South Korea, 2 Graduate School of
Information, Yonsei University, Seoul, South Korea

a1111111111 * [email protected]
a1111111111
a1111111111
a1111111111 Abstract
a1111111111
The stress placed on global power supply systems by the growing demand for electricity has
been steadily increasing in recent years. Thus, accurate forecasting of energy demand and
consumption is essential to maintain the lifestyle and economic standards of nations sus-
OPEN ACCESS tainably. However, multiple factors, including climate change, affect the energy demands of
Citation: Chung J, Jang B (2022) Accurate local, national, and global power grids. Therefore, effective analysis of multivariable data is
prediction of electricity consumption using a hybrid required for the accurate estimation of energy demand and consumption. In this context,
CNN-LSTM model based on multivariable data. some studies have suggested that LSTM and CNN models can be used to model electricity
PLoS ONE 17(11): e0278071. https://doi.org/
demand accurately. However, existing works have utilized training based on either electric-
10.1371/journal.pone.0278071
ity loads and weather observations or national metrics e.g., gross domestic product, imports,
Editor: Yogendra Arya, J.C. Bose University of
and exports. This binary segregation has degraded forecasting performance. To resolve
Science and Technology, YMCA, INDIA, INDIA
this shortcoming, we propose a CNN-LSTM model based on a multivariable augmentation
Received: June 13, 2022
approach. Based on previous studies, we adopt 1D convolution and pooling to extract undis-
Accepted: November 8, 2022 covered features from temporal sequences. LSTM outperforms RNN on vanishing gradient
Published: November 23, 2022 problems while retaining its benefits regarding time-series variables. The proposed model
Copyright: © 2022 Chung, Jang. This is an open exhibits near-perfect forecasting of electricity consumption, outperforming existing models.
access article distributed under the terms of the Further, state-level analysis and training are performed, demonstrating the utility of the pro-
Creative Commons Attribution License, which posed methodology in forecasting regional energy consumption. The proposed model out-
permits unrestricted use, distribution, and
performs other models in most areas.
reproduction in any medium, provided the original
author and source are credited.

Data Availability Statement: The data underlying

this study are available from GitHub at the
following link: https://github.com/Jwmc999/
HybridEnergyForecastingModel.

Funding: Funder: National Research Foundation of

Introduction
Korea Award number: NRF-2022R1F1A1063961 Literature review
Grant Recipient: Beakcheol Jang The funders had
no role in study design, data collection and Electricity is both a primary index representing the economic welfare of a country and the pri-
analysis, decision to publish, or preparation of the mary gear of modern convenience [1, 2]. Power outages threaten the quality of life and the
manuscript. economy of a state [3] and power shortages damage the economic growth of emerging nations
Competing interests: The authors have declared [3, 4]. Even nations with well-developed power systems may experience sudden power failures
that no competing interests exist. [3]. Climate change and adverse weather conditions such as heatwaves, tornadoes, and heavy

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 1 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

Table 1. The nomenclature.

Nomenclature
CNN Convolutional Neural Network
DL Deep Learning
EEC Electrical Energy Consumption
GDP Gross Domestic Product
LF Load Forecasting
LSTM Long-Short Term Memory
MAPE Mean Absolute Percentage Error
MLP Multilayer Perceptron
MTLF Mid-term Load Forecasting
RMSE Root Mean Squared Error
RNN Recurrent Neural Network
ReLU Rectified Linear Units
STLF Short-term Load Forecasting
TEC Total Energy Consumption
https://doi.org/10.1371/journal.pone.0278071.t001

snowfall increase the chances of electrical failure. Therefore, the accurate forecasting of elec-
tricity consumption has become more important to ensure a stable electricity supply.
Several studies have been conducted to predict energy consumption trends. Long short-
term memory (LSTM) models combined with convolutional neural networks (CNNs) have
been proposed for the hourly and daily forecast of energy demands, which outperform multi-
layer perceptron (MLP) and recurrent neural network (RNN) models [5–7]. Electricity load
forecasting (LF) studies [5, 7, 8] have utilized historical electricity load and climate data
uploaded in real time. Most energy consumption forecasting studies [9–12] have been con-
ducted considering factors such as population, import/export values, and gross domestic prod-
uct (GDP). The nomenclature utilized in this study is presented in Table 1.
Most existing studies on electrical energy consumption (EEC) and total energy consump-
tion (TEC) forecasting methods focus on macroeconomic features, including population,
GDP, and import/export values. On the other hand, existing studies on short-term load fore-
casting (STLF) and mid-term load forecasting (MTLF) consider only historical load and
weather/time information [13]. However, this binary separation of components constrains the
forecasting accuracy of energy demand and consumption—the components must be inte-
grated to improve prediction performance.
Table 2 presents an overview of existing load and energy consumption forecasting, includ-
ing their input and target variables, forecasting time intervals, and forecasting areas. In terms
of forecasting time intervals, energy forecasting can be categorized into two groups: 1)
monthly, seasonal, and annual forecasting [1, 10–12, 14–16] and 2) per-minute, hourly, and
daily forecasting [5–8, 17]. The former type considers socio-economic input variables from
GDP, population, sales index, production index, import/export, demographic, personality,
and Google Trends data. The latter type involves observations of the actual electric load, volt-
age, submetering, and weather. Further, the target areas of the two types are different. Methods
of the first group predict energy consumption at state [9, 14, 15], national [11, 12, 14, 16], and
global [10] levels whereas those in the second group [6–8, 17] predict the energy consumption
of relatively smaller areas. In [9], a neural network ensemble approach based on a novel sparse
adaboost framework and an echo state network was used to improve generalization ability and
construct nonlinear relations between electricity demand and other factors. The proposed
model was validated by using it to predict the industrial electricity consumption of Hubei

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 2 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

Table 2. Some typical types of energy forecasting techniques.

Ref. Input Variables Output variables Prediction Prediction Area
Interval
[9] Industrial sales, producer price index, export delivery value, export value, Industrial electricity Monthly Hubei, China
import value, etc. consumption
[10] Google trend data Oil consumption Monthly World
[11] GDP, population, export value, import value EEC/TEC Annual China
[14] Seasonal residential electricity consumption history EEC, Residential electricity Seasonal USA, 4 regions of China
consumption
[15] Electricity consumption history EEC Monthly Bejaia, Algeria
[12] Gross electricity generation, GDP, population, installed capacity, import EEC Annual Turkey
value, export value, total subscribership
[16] Energy consumption history, moving average, bias, momentum, rate of TEC Monthly USA
change, etc.
[8] Electricity load history, climatic data Residential electricity load Hourly An individual household in
Montreal, Canada
[17] Active power, reactive power, voltage, global intensity, etc Residential electricity 15 minutes An individual residential building
consumption in Paris, France
[7] Global active power, global reactive power, voltage, global intensity, sub Residential power Hourly An individual household in France
metering consumption
[5] Electricity load history (ISO-NE, NYISO 13 years of electricity load of Electricity load Hourly New England and New York, USA
New England, New York)
[6] Electricity load history, climatic data, time information, electricity rates, Electricity load Daily Three distribution complexes of
number of sensors South Korea
https://doi.org/10.1371/journal.pone.0278071.t002

Province of China. However, further real-life applications in different regions of China are
required to establish its generality. In [10], Google Trends data, including the search history of
certain keywords, were used as the input dataset and an online big data-driven oil consump-
tion forecasting approach was constructed for application alongside statistical and machine
learning (ML) techniques. It was the first attempt to use the trends of google search data to pre-
dict uncertain yet essential oil consumption. It also proved the predictive power of Google
search trends through relationship investigation. However, the proposed prediction improve-
ment techniques did not explore deep learning (DL) modules. In [14], the forecasting speed of
residential electricity consumption was prioritized, contrary to most residential electricity con-
sumption studies that focus on forecasting accuracy. The authors proposed a hybrid Improved
Whale Optimization Algorithm—Optimized Grey Seasonal Variation Index which exhibited
both high accuracy and fast convergence. Consideration of forecasting speed is practical
because real-life implementation requires fast forecasting. The aforementioned model also
exhibited “excellent” forecasting accuracy. In [8], the periodic part of the household residual
load forecasting was modeled based on the behavioral patterns of overestimated/underesti-
mated residual components. The model exhibited a significant improvement in the prediction
accuracy of periodic residual demand. When combined with climatic data, it also improved
total power consumption prediction. Nevertheless, as the real dataset of the experiment per-
tained to a single household, its application in broader residential households remains to be
established. To improve the performance of RNN on electric load forecasting at a specific
time, a recurrent inception convolution neural network combining RNN and 1D CNN was
proposed in [6]. To this end, a 1D CNN inception model was used to balance prediction time
and hidden state vector values. Its performance was verified based on power usage data
obtained from three distribution complexes in South Korea. The model outperformed the
benchmarks of MLP, RNN, and 1D CNN. However, because there are multitudinous power

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 3 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

distribution complexes in South Korea, its application to the entire country requires further
verification.

Research gap and motivation

In this paper, we present a novel and accurate hybrid CNN-LSTM model that uses multivari-
able data to forecast EEC. Further, we explain the potential problems and roles of various
state-level forecasting components, such as regional import and export values. The proposed
technique utilizes broad and necessary components of state-level time-series forecasting tech-
niques. The proposed technique uses LSTM layers to process and predict based on time-series
data, and the output depends not only on the current input but also on the inputs from the
past. The hybrid CNN-LSTM also utilizes convolutional and pooling layers to extract various
aspects and objects in a time series, and achieves differentiation at a relatively low computa-
tional cost. Moreover, it uses a multivariable dataset containing the data of electric load, export
values, and import values to estimate the electricity usage of the following month. Industrial
electricity usage and imports/exports are directly proportional to each other; hence, the value
of international trade represents the present and future costs of electricity closely. Our tech-
nique also allows users to analyze and predict electricity consumption of any region, in addi-
tion to that on a national level.
South Korea was deemed to be an adequate data source to identify the correlation between
import/export values, industrial electricity usage, and total electricity consumption. As an
advocate for climatic environmentalism, the South Korean government has recently reduced
the number of nuclear and thermal power plants in the country. This endeavor has led to
nationwide concerns regarding electricity shortage, as 72.9% of the GDP [18] of the country
relies on importing/exporting sectors of the manufacturing industry, which require more than
one-third of the nation’s total EEC [19]. Moreover, manufacturing and production firms in
South Korea are primarily distributed in specific regions, necessitating accurate state-level
energy forecasting for smooth operation. Therefore, 16 regions in South Korea are analyzed to
train a novel nationwide DL-based EEC forecasting model in this paper.
In this study, the Open Data Portal [20], Open MET Data Portal [21], and Trade Statistics
[22] government databases of South Korea are used, with each offering monthly data on
energy consumption, import/ export values, and temperature and rainfall data during the
period between January 2004 and December 2020. Our augmented multivariable CNN-LSTM
is observed to exhibit a root mean square error (RMSE) of 0.165 in the nation-level experi-
ment, outperforming conventional deep learning-based energy forecasting algorithms—the
RMSEs of MLP, RNN, and LSTM are 0.4521, 0.1713, and 0.174, respectively. Moreover, in
regional forecasting, the CNN-LSTM model using the proposed data processing method out-
performs the alternatives in all the areas, accurately predicting 16 states in terms of RMSE and
mean absolute percentage error (MAPE).

Contribution
As mentioned in the Literature Review and the Research Gap and Motivation sections, previ-
ous approaches to the prediction of electricity consumption and demands have not considered
the two types of input factors simultaneously. However, greater integrity in input is necessary
to improve prediction accuracy. Moreover, monthly state-level predictions are required to
account for variable energy consumption patterns of different states. Finally, we utilize an
interpolation technique for convergence between monthly data points. While the time units of
LF datasets are as detailed as seconds and milliseconds, the finest unit of time in the case of

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 4 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

EEC/TEC is a month. Our interpolation bridges these gaps between different data points and
ensures better prediction scores. The various contributions of this paper are as follows:
• We present a novel EEC forecasting method based on small datasets using data
augmentation.
• The proposed model integrates input components of EEC/TEC forecasting and LF to
improve prediction performance.
• To the best of our knowledge, this is the first study to conduct the state-level monthly EEC
forecasting for South Korea.

Paper organization
The rest of this paper is structured as follows. In Materials and Methods, the proposed deep
learning model is introduced along with the requisite background information. Experimental
validation of the proposed method is presented in the Results section. In the Discussion sec-
tion, the motivation of the study is re-stated and the results obtained using the proposed
method are analyzed. Finally, the study is concluded in the Conclusions section.

Materials and methods

The proposed energy-consumption forecasting model comprises three steps. In the first step,
the univariable and multivariable time series data based on continuous observations of clima-
tological/economic information are analyzed and a collective dataset is constructed. In the sec-
ond step, a resampling method and spline interpolation are applied to generate the simulation
dataset comprising continuous polynomial segments. An overview of deep learning technol-
ogy for sequential data is introduced in the third step with the objective of preserving an accu-
rate one-day and one-month forecasting horizon for electricity consumption.

Extraction of important features

In [1, 2], a close correlation was reported between GDP and power consumption. However,
the annuity of GDP impedes its application in deep learning, as annual GDP index data is too
sparse for DL model training. To offset this challenge, monthly macroeconomic features,
including import/export values, volume, and trade gains and losses from Korea Customs Ser-
vice, are collected in this study. Fig 1 depicts the Pearson correlation coefficients of all 11 fea-
tures. The import and export values exhibit the highest correlation with power usage. Climatic
factors, such as temperature and precipitation, exhibit weak relationships with the target
variable.
Regional analysis is another important sector of energy forecasting because energy con-
sumption is correlated to the local conditions of each community. These conditions include
weather, population, and economic factors. However, although studies have been conducted
on provincial energy forecasting [5, 9, 14], nationwide forecasting of each region remains to be
performed.

Data augmentation
Certain input variables cannot be collected at the frequency of a target variable. Specifically,
macroeconomic factors, such as import and export values, are usually collected monthly, quar-
terly, and annually, while empirical observations of temperature, rain, and electric load are

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 5 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

Fig 1. Correlation coefficients of input variables.

https://doi.org/10.1371/journal.pone.0278071.g001

measured every second, minute, or hour. To preserve the characteristics of both domains, a
new dataset of daily frequency can be used as a middle ground.
Fig 2 illustrates the workflow of the proposed method. Once the rows of new daily values
are generated, interpolation is used to generate the simulated values. In power forecasting, as
mentioned in [23], piecewise cubic polynomial interpolation represents a reasonable compro-
mise between computational cost and flexibility. In [24] as well, piecewise cubic interpolation
was preferred over other spline methods to smooth the down-sampled data, suggesting a data-
driven load-forecasting method. However, in the aforementioned study, a significant perfor-
mance improvement was not reported in the case of the cubic spline compare to the case of
quadratic spline—instead, piecewise cubic interpolation required greater computational time

Fig 2. Visualization of the proposed methodology.

https://doi.org/10.1371/journal.pone.0278071.g002

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 6 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

during the training process than the quadratic spline. Therefore, piecewise quadratic-spline
interpolation is used for data augmentation in this study.
The function, S(x), interpolates each local data point piecewise to restrict Runge’s phenom-
enon. Here,
2
Si ðxÞ ¼ ai þ bi ðx xi Þ þ ci ðx xi Þ ; ð1Þ

where i 2 [0, 1, . . ., n] and x 2 R. The parameters ai, bi, and ci denote quadratic polynomial
coefficients. A multi-dimensional array of size 6180 × lag × features defines a new daily poly-
nomial coefficient structure. For each region, the length of the time series is taken to be 6180,
lag denotes the length of the shifted time-frequency for single-step forecasting, and features
denote multivariable factors from the given datasets containing data regarding electricity con-
sumption, weather information, and import/export indexes.

Deep learning models

1. Multilayer perceptron. MLP was introduced to address the limitations of the percep-
tron. By stacking one or more hidden layers between the input and output layers, MLP maps
the input and output nonlinearly. First, the input layer receives an input tensor of arbitrary
shape. The weight of each neuron is calculated as the data transmitted from the input to the
output layer. Then, the error is minimized by backpropagating the mini-batch gradient
descent (for weight update) for every epoch. Notably, the performance of the feedforward net-
work depends on the learning rate and the effectiveness of the optimizer of the backpropaga-
tion process. Finally, the output layer of the MLP yields the desired output, e.g., the
classification or prediction result. The simplest form of MLP is based on linear regression and
comprises a single input and output layer with no activation functions (i.e., a linear activation
function). Eq (2) presents the mathematical expression of MLP, where w denotes the vector of
weights, x denotes the input tensor, b denotes the bias, and σ denotes the nonlinear activation
function.
!
Xn
y¼s wi xi þ b ¼ sðW T x þ bÞ ð2Þ
i¼1

Despite its nonlinearity, MLP experiences overfitting and vanishing gradient problems. Over-
fitted models fail to predict test data correctly, thereby diminishing their feasibility. In addi-
tion, in MLPs containing hundreds of hidden layers, the backpropagation process can
diminish the gradient to 0, or close to 0, impeding further training.
2. Recurrent neural networks. RNNs and RNN-based models exhibit superior perfor-
mance on sequence data, such as text and time series data [25]. While feedforward neural net-
works remember only the current input, RNNs store information regarding the temporal
order and consider previous inputs as well as the current state during decision-making. As in
the case of a feedforward network, an RNN first calculates the loss function on a batch (for-
ward propagation). Then, it updates the gradients based on the current state by calculating the
state memory of earlier instances (backpropagation through time).
ahti ¼ tanhðWax xhti þ Waa aht 1i
þ ba Þ ð3Þ

^y hti ¼ softmaxðWya ahti þ by Þ ð4Þ

Eq (3) presents a basic RNN cell, where xhti denotes the current input, and aht−1i denotes a
previous hidden state with previous information. A single RNN unit takes xhti and aht−1i as

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 7 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

inputs and outputs ahti to the following RNN cell. Output ahti is also used to predict yhti. How-
ever, RNNs are also prone to the vanishing gradient problem, as their cells have short-term
memory.
3. Long-short term memory. To counteract the vanishing gradient problem of RNNs,
three gates that forget, update, and output the hidden states of current and previous instances
are used in LSTM [26]. In Eq (5) the forget gate decides whether to delete or retain stored
memory values of the input state. If one of the values of Ghti f is 0, or close to 0, the LSTM cell

removes information from the corresponding component of cht−1i. If one of the values is 1, the
LSTM cell retains the information. Similarly, the update gate reflects the information from the
forget gate, whereas the output gate determines the outputs to be used. All these decisions are
then added to the cell state, and the memory variable is subsequently delivered to the subse-
quent LSTM cell with a hidden state. Thus, each LSTM-cell tracks and updates a memory vari-
able chti (i.e. a cell state) at each time step, which can be different from ahti.

Ghti
f ¼ sðWf ½a
ht 1i
; xhti � þ bf Þ ð5Þ

Ghti
u ¼ sðWu ½a
ht 1i
; xhti � þ bu Þ ð6Þ

^c hti ¼ tanhðWc ½aht 1i

; xhti � þ bc Þ ð7Þ

chti ¼ Ghti
f �c
ht 1i
þ Ghti
u �^c hti ð8Þ

Ghti
o ¼ sðWo ½a
ht 1i
; xhti � þ bo Þ ð9Þ

ahti ¼ Ghti hti

o � tanhðc Þ
ð10Þ

4. Convolutional neural network and long-short-term memory. CNN-LSTM architec-

tures extract detailed features from multiple variables to forecast EEC and can remember com-
plex irregular trends [7]. The CNN-LSTM model comprises two stages—feature extraction
and forecasting [27], as illustrated in Fig 3. Feature extraction comprises a 1D CNN consisting
of convolution layers, an activation function, and a pooling layer. The 1D CNN extracts fea-
tures from an input layer that receives multivariable macroeconomic and observational time-

Fig 3. CNN-LSTM model architecture.

https://doi.org/10.1371/journal.pone.0278071.g003

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 8 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

series sequences. Then, the CNN layer transmits the results to the LSTM layer, and the dense
layer forecasts the future electricity consumption of each region.
In Eq (11), hij1 denotes the output vector of the first convolution layer, where i 2 [0, 1, . . .,
n], xi0 ¼ x1 ; x2 ; . . . ; xn denotes the input energy consumption input vector, n denotes the
length of time-series input sequence, j denotes the index value of the feature map correspond-
ing to each lag, M denotes the number of filters, and W denotes the weight vector of the kernel.
hlij denotes the result of the lth convolution layer and the ith value of the layer. To reduce the
network computation costs and the number of trainable parameters, the CNN layer uses a
pooling layer to reduce the spatial size of the representation. The pooling layer plij integrates a
neuron filter on a previous layer into a scalar in the subsequent layer, where T denotes the size
of the stride and R denotes the pooling size. Max-pooling is used in this paper, which selects
the maximum value from each filter during feature extraction.
!
X
M
1 1 0 1
h ¼s
ij W x m;j iþm 1;j þb j ð11Þ
m¼1

!
X
M
l l 0 l
h ¼s
ij Wm;j xiþm 1;j þb j ð12Þ
m¼1

plij ¼ max hli�Tþr;j

1
ð13Þ
r2R

The kernel of the 1D CNN moves vertically across a feature map; hence, it receives a single
integer as the filter height. In contrast, a 2D CNN has a two-dimensional kernel, e.g., a 2 × 1 fil-
ter [7]. A 1D CNN is selected in this study instead of a 2D CNN because our input data con-
sists of a one-dimensional time series sequence. In addition, 1D CNN is more competitive in
terms of computational cost than 2D CNN [28].
5. Evaluation metrics. The proposed model was compared with existing ones in terms of
RMSE and MAPE. RMSE calculates prediction errors and indicates how far the residuals are
from the line of best fit. RMSE is useful to configure models for a specific variable and compare
prediction errors between different models. RMSE is defined as follows:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1X n
2
RMSE ¼ ðC At Þ ð14Þ
n t¼1 t

where n denotes the number of observations, Ct denotes the forecasted values of consumption,
and At denotes the observed actual values at time stamp t.
MAPE is used to compare the forecasting accuracy of different time-series models. For each
forecasted point in period t, the prediction error is given by et = At − Ct, and the absolute value
of the percentage error, et = (At − Ct)/At, is summed and divided by the number of fitted
points, n. MAPE corresponding to a period t is given by:
� �
1X n �
� ðAt Ct Þ��
MAPE ¼ � � 100 ð15Þ
n t¼1 � At

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 9 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

Results
Data pre-processing
For data pre-processing, the multivariable time-series data are rescaled from monthly to daily
periods. Each feature contains 204 rows of monthly values collected between January 2004 and
December 2020 (204 months). 204 months are upsampled into 6180 days for 16 South Korean
regions. The upsampled dates contain NaN values for all 3 features of historical electricity
usage, export value, and import value. Historical electricity consumption is defined as the sum
of residential, industrial, educational, and other electricity usages. Export and import values
denote the total price of exported and imported merchandise, respectively. As mentioned in
Extraction of Important Features, because the climate variable has little correlation with histor-
ical electricity usage, climatic data is excluded from experiments. The empty values are esti-
mated using spline interpolation. The total length of the preprocessed data is taken to be
6180 × 16 × 4, without any missing values. Following previous energy forecasting studies [6],
MLP is adopted as our baseline model. The baseline method is taken to be univariate single-
step forecasting based on a previous consumption time series. To confirm the effect of aug-
mentation, multivariable forecasting is also included in this experiment. The unaugmented
univariate and multivariable methods are used to predict the monthly consumption based on
the data of the previous two months. The proposed multivariable augmented model uses the
data of the interpolated past 30 days to predict the consumption figure for each day. To com-
pare the daily forecasts with the monthly forecasts, we extract consumption predictions for
specific dates corresponding to the time intervals before interpolation.

Model hyperparameters
To focus on the impact of data augmentation, the architecture of the DL models is simplified.
The MLP consists of a single dense layer with one node and one linear activation function.
Both the RNN and LSTM comprise 128 units, a ReLU (Rectified Linear Unit) activation func-
tion and a dropout layer are included following a single hidden layer. The CNN-LSTM con-
tains 2 layers of 1D CNN with 128 filters, kernel size of 1, stride of 1, pooling size of 2, and a
ReLU activation function, followed by the LSTM layer with identical hyperparameter as the
LSTM-only model. All models accept a 2D array of (length of dataset, number of features) as
the input and use the Adam optimizer for compilation. The final dense layer of the augmented
model forecasts the power consumption for the subsequent day and those of the other predicts
the consumption for the subsequent month. All models split the training and test datasets fol-
lowing a 70:30 ratio. The training is repeated for each model, region, and method to ensure a
fair comparison.

Comparison of national performance

The performance of the proposed model is compared with those of other DL algorithms in
terms of model fitness on different examples. Notably, LSTM does not always yield superior
results compared to RNN [29], even though the former improves the exploding/vanishing gra-
dient problems of RNNs. The performance of each algorithm is estimated on a basic dataset,
that is univariate and without augmentation. Fig 4 compares the national electricity usage pre-
dictions obtained using deep learning models with those of the baseline monthly dataset. This
figure confirms that the time-series outputs of the RNN, LSTM, and CNN-LSTM are almost
identical. Table 3 summarizes the evaluation metrics used for further comparison of the mod-
els. The RMSE and MAPE scores are observed to decrease significantly with the addition of
recurrent layers, and the inclusion of a 1D convolution layer is observed to improve

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 10 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

Fig 4. Comparison of national forecasting performance.

https://doi.org/10.1371/journal.pone.0278071.g004

performance additionally. Also, to further demonstrate the robustness of CNN-LSTM, we add

the forecasting comparison of the USA’s electricity consumption. Retrieved from [30], the
USA dataset consists of the national electricity usage and the price from January 1990 to July
2022. Same with the South Korea dataset, spline interpolation is applied to the monthly USA
dataset.

Regional performance comparison

We verified the effect of multivariable augmentation on the performance of the CNN-LSTM
model. The state-wise monthly electricity consumption prediction obtained using each
method on the test dataset of 60 months is illustrated in Fig 5. The proposed method effectively
minimizes the error in all regions compared to other methods, and the multivariable aug-
mented flows overlap almost perfectly with the actual power demand data. Only Gyeongnam,
Gyeongbuk, and Ulsan exhibit observable differences between the ground truth and multivari-
able augmentation. Table 4 summarizes the evaluation results of the CNN-LSTM model for 16

Table 3. Forecasting performance evaluation.

South Korea MLP RNN LSTM CNN-LSTM
RMSE 0.4521 0.1798 0.174 0.165
MAPE 0.4727 0.1602 0.1589 0.1441
USA MLP RNN LSTM CNN-LSTM
RMSE 1.9817 0.1511 0.0889 0.0761
MAPE 1.8008 0.1068 0.0574 0.0541
https://doi.org/10.1371/journal.pone.0278071.t003

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 11 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

Fig 5. State-level forecasting performance comparison of CNN-LSTM.

https://doi.org/10.1371/journal.pone.0278071.g005

regions using different data processing methods. The proposed augmented multivariable
model is observed to yield useful results in all states, exhibiting the lowest RMSE and MAPE
scores.

Ablation studies
To establish the superiority of the proposed model, two experiments are conducted as ablation
studies. First, following the 1D CNN kernel visualization technique presented in [7, 31], Fig 6
illustrates the noise-reducing power of each 1D CNN layer. The intermediate output obtained
from the second convolution network is observed to be less spiky than the first kernel. This is
because each CNN filter reduces the noise of the input dataset. This selectivity over input
information also allows cost-effective computation and is, hence, useful for mid-term time
series forecasting. The illustration aids the analysis of the number of layers that benefit from
smoothing the noise of raw data and selectively extracting important information. Moreover,
the loss landscapes of LSTM and the proposed model are compared in Fig 7. In [32], the loss
landscape was introduced as a method to represent loss convergence of a model throughout
the training process. From the figure, it is evident that the proposed CNN-LSTM hybrid
model converges the loss to the minima satisfactorily, and the graph is smooth and concave,

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 12 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

Table 4. Evaluation of forecasting performance of CNN-LSTM.

Regions RMSE MAPE
Univariable Multivariable Augmented Augmented Univariable Multivariable Augmented Augmented
Univariable Multivariable Univariable Multivariable
Busan 0.2345 0.1722 0.0853 0.0132 0.2453 0.2034 0.0935 0.0154
Chungbuk 0.3649 0.1252 0.1418 0.0253 0.3276 0.1038 0.1262 0.0193
Chungnam 0.3738 0.1108 0.0422 0.0291 0.3767 0.0970 0.0373 0.0235
Daegu 0.2456 0.2053 0.0982 0.0168 0.2937 0.2821 0.1193 0.0216
Daejeon 0.2543 0.2082 0.0782 0.0157 0.2638 0.2496 0.0862 0.0179
Gangwon 0.2038 0.1509 0.0957 0.0185 0.2107 0.1872 0.0981 0.0212
Gwangju 0.2817 0.2014 0.0988 0.0187 0.2535 0.1959 0.0939 0.0191
Gyeongbuk 0.1820 0.1416 0.0874 0.0496 0.2151 0.1750 0.1037 0.0551
Gyeonggi 0.4121 0.1662 0.1525 0.0201 0.3754 0.1415 0.1324 0.0173
Gyeongnam 0.2599 0.1883 0.0735 0.0403 0.2870 0.1983 0.0782 0.0419
Incheon 0.3165 0.1922 0.1141 0.0209 0.3168 0.2188 0.1133 0.0213
Jeju 0.4632 0.2161 0.1210 0.0273 0.3834 0.1650 0.0940 0.0210
Jeonbuk 0.3420 0.1183 0.0799 0.0232 0.4185 0.1327 0.0858 0.0253
Jeonnam 0.3344 0.1368 0.0744 0.0247 0.3849 0.1360 0.0744 0.0248
Seoul 0.2959 0.2519 0.1223 0.0170 0.4683 0.4666 0.1890 0.0274
Ulsan 0.3446 0.1580 0.0828 0.0379 0.2837 0.1085 0.0628 0.0290
https://doi.org/10.1371/journal.pone.0278071.t004

indicating facile training. On the other hand, the LSTM-only model exhibits convex loss sur-
faces and the minima are not visible, indicating that the training of the model is more difficult.
This proves that the inclusion of 1D CNN layers before the LSTM layer simplifies training.

Discussion
Electricity is the backbone of modern society. As such, the accurate prediction of electricity
demand and consumption is more significant now than ever before. However, many factors
complicate the forecasting accuracy of energy consumption, necessitating the development of
advanced forecasting models. In this paper, we presented a new hybrid CNN-LSTM multivari-
able EEC forecasting techniques that integrates the advantages of CNNs and LSTMs. To incor-
porate economic insights into energy usage, the proposed technique was expanded using
import/export values. The proposed approach enables the prediction of regional monthly elec-
tricity consumption. We tested the proposed techniques (components and algorithm) using
baseline techniques on univariable historical energy consumption data and compared them to
conventional deep learning models. Comprehensive experiment results proved that the pro-
posed technique extracts useful information between international trade values and electricity

Fig 6. Kernel output comparison of 1D CNN layers.

https://doi.org/10.1371/journal.pone.0278071.g006

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 13 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

Fig 7. Comparison of loss landscapes of LSTM and CNN-LSTM.

https://doi.org/10.1371/journal.pone.0278071.g007

usage, thereby improving the accuracy of national/regional predictions. We also expanded the
proposed technique to consider data interpolation. The proposed mechanism can be further
improved using data augmentation.

Conclusions
In this study, we investigated the effects of CNN-LSTM on augmented multivariable time
series datasets. We concluded that dimension reduction using the pooling layer of 1D CNN
reduces noise and thereby reduces the RMSE and MAPE scores. The LSTM layer was also
observed to be well suited to process time series data as it receives inputs for each time step.
Extensive experiments and ablation studies were performed, establishing the benefits afforded
by the proposed CNN-LSTM architecture paired with multivariable augmentation to provin-
cial time series forecasting for EEC.

Author Contributions
Conceptualization: Jaewon Chung, Beakcheol Jang.
Data curation: Jaewon Chung.
Formal analysis: Jaewon Chung.
Funding acquisition: Beakcheol Jang.
Supervision: Beakcheol Jang.
Visualization: Jaewon Chung.
Writing – original draft: Jaewon Chung.
Writing – review & editing: Jaewon Chung.

References
1. Wang Q, Su M, Li R, Ponce P. The effects of energy prices, urbanization and economic growth on
energy consumption per capita in 186 countries. Journal of cleaner production. 2019; 225:1017–1032.
https://doi.org/10.1016/j.jclepro.2019.04.008
2. Arora V, Lieskovsky J. Electricity use as an indicator of US economic activity; 2016.
3. Haes Alhelou H, Hamedani-Golshan ME, Njenda TC, Siano P. A survey on power system blackout and
cascading events: Research motivations and challenges. Energies. 2019; 12(4):682. https://doi.org/10.
3390/en12040682
4. Cissokho L, Seck A. Electric power outages and the productivity of small and medium enterprises in
Senegal. Investment climate and business environment research fund Report. 2013; 77:13.

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 14 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

5. Mujeeb S, Javaid N, Ilahi M, Wadud Z, Ishmanov F, Afzal MK. Deep long short-term memory: A new
price and load forecasting scheme for big data in smart cities. Sustainability. 2019; 11(4):987. https://
doi.org/10.3390/su11040987
6. Kim J, Moon J, Hwang E, Kang P. Recurrent inception convolution neural network for multi short-term
load forecasting. Energy and buildings. 2019; 194:328–341. https://doi.org/10.1016/j.enbuild.2019.04.
034
7. Kim TY, Cho SB. Predicting residential energy consumption using CNN-LSTM neural networks.
Energy. 2019; 182:72–81. https://doi.org/10.1016/j.energy.2019.05.230
8. Amara F, Agbossou K, Dubé Y, Kelouwani S, Cardenas A, Hosseini SS. A residual load modeling
approach for household short-term load forecasting application. Energy and Buildings. 2019; 187:132–
143. https://doi.org/10.1016/j.enbuild.2019.01.009
9. Wang L, Lv SX, Zeng YR. Effective sparse adaboost method with ESN and FOA for industrial electricity
consumption forecasting in China. Energy. 2018; 155:1013–1031. https://doi.org/10.1016/j.energy.
2018.04.175
10. Yu L, Zhao Y, Tang L, Yang Z. Online big data-driven oil consumption forecasting with Google trends.
International Journal of Forecasting. 2019; 35(1):213–223. https://doi.org/10.1016/j.ijforecast.2017.11.
005
11. Hu H, Wang L, Peng L, Zeng YR. Effective energy consumption forecasting using enhanced bagged
echo state network. Energy. 2020; 193:116778. https://doi.org/10.1016/j.energy.2019.116778
12. Kaytez F. A hybrid approach based on autoregressive integrated moving average and least-square sup-
port vector machine for long-term forecasting of net electricity consumption. Energy. 2020; 197:117200.
https://doi.org/10.1016/j.energy.2020.117200
13. Aslam S, Herodotou H, Mohsin SM, Javaid N, Ashraf N, Aslam S. A survey on deep learning methods
for power load and renewable energy forecasting in smart microgrids. Renewable and Sustainable
Energy Reviews. 2021; 144:110992. https://doi.org/10.1016/j.rser.2021.110992
14. Xiong Xin, Hu Xi, Guo Huan. A hybrid optimized grey seasonal variation index model improved by
whale optimization algorithm for forecasting the residential electricity consumption. Energy. 2021
https://doi.org/10.1016/j.energy.2021.121127
15. Hadjout D, Torres JF, Troncoso A, Sebaa A, Martı́nez-Álvarez F. Electricity consumption forecasting
based on ensemble deep learning with application to the Algerian market. Energy. 2022 https://doi.org/
10.1016/j.energy.2021.123060
16. Jana RK, Ghosh I, Sanyal MK. A granular deep learning approach for predicting energy consumption.
Applied Soft Computing. 2020; 89:106091. https://doi.org/10.1016/j.asoc.2020.106091
17. Liu C, Sun B, Zhang C, Li F. A hybrid prediction model for residential electricity consumption using holt-
winters and extreme learning machine. Applied energy. 2020; 275:115383. https://doi.org/10.1016/j.
apenergy.2020.115383
18. K Indicator. Exports and Imports Ratio (GDP); 2022.
19. KOSIS. Electric Power Consumption by Use; 2022.
20. OPEN DATA PORTAL. KEPCO Electricity Usage Status by Region Application; 2022.
21. Open MET Data Portal. Statistic Division; 2022.
22. Korea Customs Service. Trade Statistics for Import/Export by Region; 2022.
23. Wang G, Jia L. Short-term wind power forecasting based on BOMLS K-means similar hours Clustering
method. In: 2019 IEEE PES Asia-Pacific Power and Energy Engineering Conference (APPEEC). IEEE;
2019. p. 1–5.
24. Williams S, Short M. Electricity demand forecasting for decentralised energy management. Energy and
Built Environment. 2020; 1(2):178–186. https://doi.org/10.1016/j.enbenv.2020.01.001
25. Lipton ZC, Berkowitz J, Elkan C. A critical review of recurrent neural networks for sequence learning.
arXiv preprint arXiv:150600019. 2015.
26. Hochreiter S, Schmidhuber J. Long short-term memory. Neural computation. 1997; 9(8):1735–1780.
https://doi.org/10.1162/neco.1997.9.8.1735 PMID: 9377276
27. Livieris IE, Pintelas E, Pintelas P. A CNN–LSTM model for gold price time-series forecasting. Neural
computing and applications. 2020; 32(23):17351–17360. https://doi.org/10.1007/s00521-020-04867-x
28. Kiranyaz S, Avci O, Abdeljaber O, Ince T, Gabbouj M, Inman DJ. 1D convolutional neural networks and
applications: A survey. Mechanical systems and signal processing. 2021; 151:107398. https://doi.org/
10.1016/j.ymssp.2020.107398
29. Xiao D, Huang Y, Wang H, Shi H, Liu C. Health assessment for piston pump using LSTM neural net-
work. In: 2018 International Conference on Sensing, Diagnostics, Prognostics, and Control (SDPC).
IEEE; 2018. p. 131–137.

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 15 / 16

PLOS ONE Electricity consumption analysis in 16 states of South Korea

30. EIA. The U.S.A Energy Information Administration; 2022.

31. Tae-Young Kim, Sung-Bae Cho. Web traffic anomaly detection using C-LSTM neural networks. Expert
Systems with Applications. 2018.
32. Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, Tom Goldstein. Visualizing the loss landscape of
neural nets. Advances in neural information processing systems. 2018.

PLOS ONE | https://doi.org/10.1371/journal.pone.0278071 November 23, 2022 16 / 16