Professional Documents
Culture Documents
Park 2021
Park 2021
Park 2021
1007/s12206-021-0140-0
1. Introduction
According to the World Energy Outlook 2018 published by the International Energy Agency,
global energy demand is expected to increase by approximately 25 % by 2040, and, if energy
efficiency is not improved, the global demand will rise by approximately 50 % over the same
period [1]. Accordingly, several countries have set energy conservation and efficiency im-
provement as core strategies to address climate change and energy security by establishing
and promoting various countermeasures [2]. In addition, the global expansion of smart cities,
which intend to use information and communication technology as a new growth engine, is
rapidly progressing in accordance with the 4th industrial revolution fields such as big data
analysis, artificial intelligence, and robotics [3]. A smart city is a model that can solve urban
problems and achieve sustainability by incorporating new technologies into urban platforms
that address transportation, energy, and environmental issues [4]. At present, the Korean gov-
ernment is working to promote the greening of energy suppliers, for instance, by announcing
plans to expand the share of renewable energy generation, such as solar, geothermal, and fuel
cells, to 20 % by 2030 [5]. As part of these efforts, the implementation of zero-energy buildings
© The Korean Society of Mechanical
Engineers and Springer-Verlag GmbH and smart zero-energy cities has accelerated, and the share of renewable energy has in-
Germany, part of Springer Nature 2021 creased by 68 % [6].
795
Journal of Mechanical Science and Technology 35 (2) 2021 DOI 10.1007/s12206-021-0140-0
Therefore, advanced information technologies such as deep eration prediction results demonstrated that the proposed ap-
learning and big data have been actively applied in the building proach outperforms the persistence and SVM models and can
energy management system to improve energy efficiency. A have widespread applicability in limited data samples. Wang et
variety of studies have been performed thus far on the predic- al. [13] proposed a hybrid model based on convolutional neural
tion of renewable energy sources using machine learning network and LSTM network models, which were applied to the
technique. One such technology is recurrent neural network obtained data in the PV system. The results of the deep learn-
(RNN), an artificial neural network in which the connection ing model proposed on PV power prediction also indicate that
between units is composed of a cyclic structure that reflects the deep learning is useful in improving the accuracy of PV power
characteristics of time series. The RNN has a disadvantage in prediction.
that when the distance between the related information and the Shin and Kim [14] proposed a short-term deep learning pre-
point using the information is far, the gradient gradually de- diction model for solar power generation using meteorological
creases and the learning ability is degraded [7]. To compen- data from the Mokpo meteorological agency and generation
sate for this limitation, recent studies have explored the use of data from the Yeongam solar power plant. The forecasted
long short-term memory (LSTM) technique of RNN. The re- power generation of the proposed model show an average
search directions related to RNN and LSTM techniques in the RMSE and mean absolute error of 0.177 and 0.095 for deep
building and urban energy field are the following. neural network and 0.116 and 0.067 for RNN. Furthermore,
In a study by Jeon et aI. [7], unpredictable occupancy infor- LSTM produced the best results at 0.100 and 0.054. In another
mation was fed to a deep-learning model by assuming that, in study by Wang et al. [15], a hybrid deep learning model
the future, the occupants may actively interact with the control (LSTM-convolutional network) is proposed and applied to PV
systems through various smart devices. Deep-learning models power prediction. The results show that the hybrid prediction
with single and deeper layers were tested in this study, and model has a better prediction effect than the single prediction
both showed excellent performance for data matching during model and that the proposed hybrid model, which first extracts
the learning periods. Wen et al. [8] developed a deep RNN with the temporal characteristics of the data and then the spatial
LSTM units (DRNN-LSTM) model to forecast aggregated characteristics, is more effective than the LSTM-convolutional
power load and photovoltaic (PV) power output in a community network, which first extracts the spatial characteristics and then
microgrid. Two real-world datasets were used to test the pro- the temporal characteristics of the data. Su et al. [16] proposed
posed forecasting model, and the results show that the DRNN- a robust hybrid hours-ahead gas consumption method by inte-
LSTM model performs better than a multi-layer perception grating a wavelet transform, RNN-structured deep learning,
(MLP) network and support vector machine (SVM). The study and a genetic algorithm. The wavelet transform is used to re-
found that the energy storage system and coordinated charg- duce the complexity of the forecasting tasks by decomposing
ing mode of electric vehicles can promote peak load shifting the original series of gas loads into several sub-components.
and reduce daily costs by 8.97. Jullian et al. [9] trained deep Gensler et al. [17] introduced powerful algorithms in renewable
learning models to predict the performance of a solar hot water energy power forecasting by using different deep learning
system under different meteorological conditions. Techniques methods, such as deep belief networks, an autoencoder, and
such as artificial neural networks (ANN), RNN, and LSTM were LSTM. In their experiments, combinations of these algorithms
explored. The LSTM models achieved a low mean absolute were used to show their forecasting strength compared with a
error of 0.55 °C and the lowest root mean square error scores standard MLP and a physical forecasting model in predicting
(1.27 °C) for temperature sequence predictions, as well as the the energy output of 21 solar power plants.
lowest variance (0.52 °C) and relative prediction errors Qing and Niu [18] proposed a novel solar prediction scheme
(3.45 %) for single value predictions, indicating a more reliable for hourly day-ahead solar irradiance prediction by using
prediction performance. weather forecasting data. For a case using 10 years of histori-
Ghimire et al. [10] ascertained that a proposed hybrid model cal data to predict a year of irradiance data, the prediction
based on a convolutional network framework can accurately RMSE using the proposed LSTM algorithm decreased by
predict global solar radiation and enable energy availability to 42.9 % against back-propagation neural network. Srivastava
be regularly monitored over multi-step horizons when coupled and Lessmann [19] also investigated the LSTM technique,
with a low-latency LSTM network. Gao et al. [11] proposed comparing its forecasting accuracy to alternative methods with
day-ahead power output time-series forecasting methods in a proven track record in solar energy forecasting. From an
which ideal and non-ideal weather types were separately dis- academic point of view, LSTM and the proposed framework for
cussed. A prediction performance comparison between the experimental design provided a valuable environment for future
proposed methods with traditional algorithms revealed that the studies assessing new forecasting technology. Salman et al.
RMSE accuracy of forecasting methods based on LSTM net- [20] proposed a robust and adaptive statistical model and ex-
works can reach 4.62 % in ideal weather conditions. Han et al. plored the effect of intermediate weather variables related to
[12] proposed a method for mid-to long-term wind and PV accuracy prediction using a single-layer LSTM and multi-layer
power generation prediction based on a copula function and LSTM models. The best LSTM model was a multi-layer LSTM,
LSTM network. The study’s independent and joint power gen- and the best intermediate data were pressure variable. Based
796
Journal of Mechanical Science and Technology 35 (2) 2021 DOI 10.1007/s12206-021-0140-0
797
Journal of Mechanical Science and Technology 35 (2) 2021 DOI 10.1007/s12206-021-0140-0
can also add or remove information in the cell state, which is Table 1. Cv(RMSE) value (%) according to number of hidden neurons and
controlled by gates [27]. In the first step, the forget gate deter- hidden layers with a learning rate of 0.05.
mines what data are discarded from the cell state. Then, the Number of hidden layers
input gate determines the data to be added and stores it in the Sort
1 2 3 4
cell state. Finally, the output gate is controlled by determining
10 14.89 13.92 13.79 13.96
the output data and propagating them to the next node. There-
Number 100 15.95 13.54 13.96 14.53
fore, the LSTM technique is suitable for the continuous data
of hidden 250 16.25 13.77 13.08 13.92
processing in the present simulation.
neurons
300 16.11 13.8 13.65 14.45
3. Predictive model based on deep learn- 500 13.8 13.65 13.28 50.45
ing
3.1 Learning process Table 2. Cv(RMSE) value (%) according to number of hidden neurons and
hidden layers with a learning rate of 0.01.
In this study, we attempted to implement an RNN, an ANN
Number of hidden layers
capable of continuous input processing, to predict solar PV Sort
1 2 3 4
power generation. The RNN model consisted of various learn-
ing variables, such as learning rate, number of layers, number 10 15.83 14.53 50.45 50.54
of units, and optimization techniques that constitute LSTM [7]. Number 100 15.64 13.89 13.75 50.54
This study developed its models using the LSTM toolbox pro- of hidden 250 15.68 13.61 13.71 14.21
vided by MATLAB. MATLAB recommends using the stochastic neurons
300 15.89 14.14 13.54 14.38
gradient descent (SGD) method and Adam algorithms as the 500 15.96 13.92 14.34 13.48
optimization method of LSTM for RNN implementation [25].
The SGD technique is considered inefficient because it takes a Table 3. Cv(RMSE) value (%) according to number of hidden neurons and
significant amount of time to find an optimal solution when the hidden layers with a learning rate of 0.005.
working environment is insufficient for iterative calculation [7].
Number of hidden layers
Therefore, in this study, to repeatedly predict the amount of Sort
solar PV power generation, we used the Adam algorithm, 1 2 3 4
which can find the optimal solution efficiently by flexibly adjust- 10 15.01 13.91 50.45 50.45
ing the learning rate and model building time through learning. Number 100 15.62 14.24 50.45 50.45
Then, various training options work as variables that influence of hidden 250 15.78 14.07 50.45 13.75
learning. We set the number of hidden layers and hidden neu- neurons
300 15.71 14.75 14.17 50.45
rons by which the predictive model could derive the most accu- 500 15.23 13.83 14.25 13.48
rate results.
The learning rate is an additional factor that directly affects Table 4. Cv(RMSE) value (%) according to number of hidden neurons and
the prediction results. If the learning rate is extremely small or hidden layers with a learning rate of 0.001.
large, learning can take a long time or expand considerably.
Therefore, an important task is to find the suitable learning rate Number of hidden layers
Sort
for the predictive model. Predictive values were calculated 1 2 3 4
through the aforementioned learning variables by applying the 10 14.72 15.33 50.45 50.45
coefficient of variation of the root mean squared error, Number 100 15.18 14.87 50.45 50.45
Cv(RMSE), which is a statistical concept for determining overall of hidden 250 15.18 13.59 50.45 50.45
accuracy; the optimal LSTM model was set by adopting the neurons
300 15.31 13.71 50.45 50.45
variable with the lowest Cv(RMSE) value [27, 28]. Although the 500 15.32 13.96 13.87 50.45
model shows the prediction accuracy of more than a certain
level within the training dataset, when applied to new data, what
does not fit well is called overfitting. To prevent this condition, RMSE period
CV ( RMSE ) = (2)
we divided the learning and test sets. The simulation was per- Aperiod
formed by classifying 80 % of the total data as learning data and
20 % as testing data. In addition, the model was normalized for Aperiod =
∑ period
M interval
. (3)
accurate prediction, preventing the distortion or diversion of N interval
learning results through normalization of input and output values.
3.2 Development of predictive model
∑(S − M )
2
RMSE = interval
(1) Tables 1-4 show the Cv(RMSE) values according to the
N interval learning rate. Four learning rates were set: 0.05, 0.01, 0.005,
798
Journal of Mechanical Science and Technology 35 (2) 2021 DOI 10.1007/s12206-021-0140-0
799
Journal of Mechanical Science and Technology 35 (2) 2021 DOI 10.1007/s12206-021-0140-0
800
Journal of Mechanical Science and Technology 35 (2) 2021 DOI 10.1007/s12206-021-0140-0
tion value of the single LSTM model was 157.3 kWh, and the
prediction value of the multiple LSTM model was 147.2 kWh.
Although a slight difference was observed between the meas-
ured and predicted values, the error rate rarely occurred in the
last two days. On the other hand, the maximum amount of
solar radiation was 0.96 kWh. During the same 4-day period,
the measured value of solar PV power generation was
Fig. 11. Solar PV power generation and solar radiation in a 4-day period.
488 kWh, the prediction value of the single-LSTM model was
484.4 kWh and the prediction value of the multi-LSTM model
curate prediction is possible. was 468 kWh. An accurate prediction was made, with a mini-
Fig. 10 shows some of the learning performance that pref- mal error rate between the measured and predicted values.
aced the predictions. Six days of data training were performed Finally, with regard to the averages of solar radiation and solar
and one-day testing was presented. During the learning period, PV power generation, the average value of solar radiation over
the measured and predicted graphs moved in the same pat- the 4 days was 0.59 kWh. The measured solar PV power gen-
tern; however, a slight difference was observed in the amount eration was 303.3 kWh, the single LSTM prediction was
of power generation. The difference appears to be largely due 285.1 kWh and the multiple LSTM prediction was 286.2 kWh.
to the lack of learning data, and if the number of learning data This mean value was similar for the measured and predicted
is increased, the errors should be reduced in both the learning values.
and prediction periods. Furthermore, the difference was im- Fig. 12 shows a comparison of the three graphs: the actually
proved during the prediction (testing) period of one day be- measured solar PV power generation and solar PV power
cause the testing was performed after the data training for six generation predicted by the single-LSTM and multi-LSTM
days. The multiple-layer LSTM model also changes in a pattern models. The graph shows that the measured and predicted
more similar to the measured data than does the single-layer graphs move in a similar pattern, although errors are found at
LSTM model. some intervals. The single-LSTM model and multi-LSTM mod-
Fig. 11 displays the predictions of solar PV power generation els moved in almost the same pattern. A Cv(RMSE) of the
from 09:00 to 17:00 for 4 days after data training and the single LSTM showed an error rate of 13.8 %, while the
amount of solar radiation for 4 days. The bar graph indicates Cv(RMSE) of the multiple LSTM showed a 13.2 % error rate.
the amount of solar radiation during that period. The solar ra- At high solar PV power generation, the multi-LSTM model was
diation strengthens in the daytime as the time changes and the able to predict more accurately than the single-LSTM model.
solar PV power generation moves in the same pattern as the Although no noticeable difference was found, the multi-LSTM
solar radiation does. The minimum solar radiation level over 4 model showed a slightly improved Cv(RMSE) and a prediction
days is 0.18 kWh. During the same 4-day period, the meas- closer to the measured value. Table 5 summarizes the learning
ured amount of PV power generation was 80 kWh, the predic- performance of the two LSTM models.
801
Journal of Mechanical Science and Technology 35 (2) 2021 DOI 10.1007/s12206-021-0140-0
802
Journal of Mechanical Science and Technology 35 (2) 2021 DOI 10.1007/s12206-021-0140-0
to-long term wind and photovoltaic power generation prediction adaptive control of a VAV-HVAC&R system, Int. J. Air-Cond.
based on copula function and long short term memory network, Ref., 27 (2019) 1950006.
Applied Energy, 239 (2019) 181-191. [23] S. Li and M. Zaheeruddin, A model and multi-mode control of
[13] K. Wang, X. Qi and H. Liu, A comparison of day-ahead a centrifugal chiller system: a computer simulation study, Int. J.
photovoltaic power forecasting models based on deep learning Air-Cond. Ref., 27 (2019) 1950031.
neural network, Applied Energy, 251 (2019) 113315. [24] M. P. Swarup and A. P. Kumar, Effective implementation of
[14] D. H. Shin and C. B. Kim, Short term forecast model for solar value engineering using artificial neural network aid of optimi-
power generation using RNN-LSTM, The Journal of Korea zation techniques, Int. J. Air-Cond. Ref., 27 (2019) 1950022.
Navigation Institute, 22 (3) (2018) 233-239. [25] MathWorks, http://www.mathworks.com (2019).
[15] K. Wang, X. Qi and H. Liu, Photovoltaic power forecasting [26] T. H. Jo, Deep-learning for Everyone, Gilbut (2019) 234-238.
based LSTM-convolutional network, Energy, 189 (2019) [27] J. M. Lee, S. H. Hong, B. M. Seo and K. H. Lee, Application of
116225. artificial neural networks for optimized AHU discharge air tem-
[16] H. Su, E. Zio, J. Zhang, M. Xu, X. Li and Z. Zhang, A hybrid perature set-point and minimized cooling energy in VAV sys-
hourly natural gas demand forecasting method based on the tem, Applied Thermal Engineering, 153 (2019) 726-738.
integration of wavelet transform and enhanced deep-RNN [28] S. Yeon, B. Yu, B. Seo, Y. Yoon and K. H. Lee, ANN based
model, Energy, 178 (2019) 585-597. automatic slat angle control of venetian blind for minimized to-
[17] A. Gensler, J. Henze, B. Sick and N. Raabe, Deep learning tal load in an office building, Solar Energy, 180 (2019) 133-145.
for solar power forecasting — an approach using autoencoder
and LSTM neural networks, IEEE International Conference on
Systems, Man, and Cybernetics (SMC) (2016) 002858-002865. Minkyeong Park is pursuing a Master’s
[18] X. Qing and Y. Niu, Hourly day-ahead solar irradiance predic- degree at the Department of Architecture,
tion using weather forecasts by LSTM, Energy, 148 (2018) College of Engineering, Korea University,
461-468. Seoul, Korea.
[19] S. Srivastava and S. Lessmann, A comparative study of
LSTM neural networks in forecasting day-ahead global hori-
zontal irradiance with satellite data, Solar Energy, 162 (2018)
232-247.
[20] A. G. Salman, Y. Heryadi, E. Abdurahman and W. Suparta,
Single layer & multi-layer long short-term memory (LSTM) Kwang Ho Lee is an Associate Profes-
model with intermediate variables for weather forecasting, Pro- sor of the Department of Architecture,
cedia Computer Science, 135(2018) 89-98. College of Engineering, Korea University,
[21] W. H. Kang, J. M. Lee, S. H. Yeon, M. K. Park, C. H. Kim, J. Seoul, Korea. He received his Ph.D.
H. Lee, J. W. Moon and K. H. Lee, Modeling, calibration, and from the School of Architecture from the
sensitivity analysis of direct expansion AHU-water source VRF University of Illinois at Urbana-Cham-
system, Energy, 199 (2020) 117435. paign. His research interests include
[22] M. Ning and M. Zaheeruddin, Neural network model-based advanced HVAC system control.
803