Deep Learning Based Techniques to Enhance the Performance of Microgrids: A Review

Abstract

In the last few years, carbon emissions and energy demand have increased dramatically around the globe due to a surge in population and energy-consuming devices. The integration of renewable energy resources (RERs) in a power supply system provides an efficient solution in terms of low energy cost with lower carbon emissions. However, renewable sources like solar panels have irregular nature of power generation because of their dependence on weather conditions, such as solar radiation, humidity, and temperature. Therefore, to tackle this intermittent nature of solar energy, power prediction is necessary for efficient energy management. Deep learning and machine learning-based methods have frequently been implemented for energy forecasting in the literature. The current work summarizes the state-of-the-art deep learning-based methods that are proposed to forecast the solar power for proper energy management. We also explain the methodologies of solar energy forecasting along with their outcomes. At the end, future challenges and opportunities are uncovered in the application of deep and machine learning in this area.

Full text

Deep Learning based Techniques to Enhance the
Performance of Microgrids: A Review
Sheraz Aslam1,2, Herodotos Herodotou1, Nasir Ayub3, and Syed Muhammad Mohsin2
1Department of Electrical Engineering, Computer Engineering and Informatics,
Cyprus University of Technology, Limassol 3025, Cyprus
2Department of Computer Science, COMSATS University Islamabad, Islamabad 45550, Pakistan
3Federal Urdu University of Arts, Science and Technology, Islamabad 44000, Pakistan
Correspondence: sheraz.aslam@ieee.org, herodotos.herodotou@cut.ac.cy
Abstract—In the last few years, carbon emissions and energy
demand have increased dramatically around the globe due to a
surge in population and energy-consuming devices. The integra-
tion of renewable energy resources (RERs) in a power supply
system provides an efficient solution in terms of low energy cost
with lower carbon emissions. However, renewable sources like
solar panels have irregular nature of power generation because of
their dependence on weather conditions, such as solar radiation,
humidity, and temperature. Therefore, to tackle this intermittent
nature of solar energy, power prediction is necessary for efficient
energy management. Deep learning and machine learning-based
methods have frequently been implemented for energy forecasting
in the literature. The current work summarizes the state-of-the-
art deep learning-based methods that are proposed to forecast
the solar power for proper energy management. We also explain
the methodologies of solar energy forecasting along with their
outcomes. At the end, future challenges and opportunities are
uncovered in the application of deep and machine learning in
this area.
Index Terms—Deep learning based techniques; Forecasting;
Artificial neural network; Machine learning; Weather forecast-
ing; Energy forecasting; Renewable energy resources
I. INTRODUCTION
The world is moving towards renewable energy resources
(RERs) due to their low cost and huge contributions in alle-
viation of carbon emissions. RERs consist of various sources,
including wind energy, bioenergy, hydropower, solar energy,
etc., and usually, these sources operate in islanded or grid-
connected modes. Fossil fuels-based energy sources are used
in many countries to meet the power demand of the consumers;
however, these resources are inefficient and inadequate [1].
Solar energy is generated through installing solar panels and
it is available in most locations of the world. Moreover, solar
energy plays a significant role in green energy among all
RERs [2], [3]. Figure 1 shows yearly percentage of renewable
energy contribution in total energy generation of some leading
countries around the globe. This figure shows that Brazil is
generating the highest renewable energy around the globe for
meeting the power demand of the consumers.
Solar panel converts direct sunlight to electrical energy. One
of its key characteristics is its intermittent and unpredictable
nature as the energy generated from solar panels totally
depends on environmental conditions, like temperature, solar
radiation, etc. For instance, solar panels generate maximum
electricity when the sun has high radiations (clear sky); in
contrary, it generates minimum power (maybe zero) in night
time or durng cloudy weather. However, a huge amount of
fluctuation in energy generation from solar panels may create
problems in power systems, including voltage irregulations,
power distribution, and reserve power flow problems. Unde-
sired voltage fluctuation is considered the main problem in
power distribution due to solar panels and it may lead to
instability of a microgrid [4].
1990 1995 2000 2005 2010 2015 2020
Years
0
10
20
30
40
50
60
70
80
90
100
RE contribution (%)
World
European Union
Brazil
Colombia
United Kingdom
Canada
America
Asia
Sweden
Fig. 1: Renewable energy contribution in total energy produc-
tion around the globe [5]
To ensure the safe operation of the power system and the
balance between energy demand and supply, accurate forecast-
ing of power generation from solar panels is an exigent need.
Such accurate forecasting enables the power utility operator
to efficiently manage supply with demand and generate excess
electricity from brown energy sources in case of less energy
generation from renewable sources. The forecasting of solar
energy generation is a complex task because it completely
depends on weather conditions (e.g., temperature, humidity,
solar irradiance, and cloud) [6], [7], [8]. Various forecasting
models are used to predict solar energy in the literature and

Microgrid
Storage systems
Energy Market
Smart Transportation
Mobile Applications
External Grid
Smart Home
Renewable Sources
Weather Forecast
Communication Flow Power Flow
Fig. 2: A typical microgrid architecture [16]
these models are categorized into physical, statistical, and
empirical models.
Accurate energy forecasting is essential for effective energy
planning and management. A lot of forecasting methods are
presented in the literature. For instance, [9] proposed a solar
energy forecasting method using deep learning, [10] presentd
a short-term solar energy forecasting method using statistical
methods, while a recurrent neural network (RNN) model is
used in [11]. Chen et al. developed a solar energy prediction
model that is based on convolutional neural networks (CNN)
[12] and the authors of [15] have developed a deep CNN
based model that is used for solar energy forecasting. This
paper presents a comprehensive review of the deep learning-
based techniques that are used in the literature to predict solar
energy. We have explained some key deep learning techniques
in the Section II, while Section III presents deep learning
in energy management systems (EMSs) and different solar
energy forecasting models. Section IV uncovers the future
challenges to current methods and opportunities. Finally, the
last section concludes the study.
II. DE EP LE AR NI NG BA SE D MET HO DS
This section presents the most commonly used deep learning
based methods for energy management and power forecast-
ing, namely, artificial neural networks (ANN), deep neural
networks (DNN), convolutional neural networks (CNN), and
recurrent neural networks (RNN).
A. Artificial Neural Networks
An artificial neural network (ANN) is designed on basis of
working mechanism of the human nervous system [17]. Such
a system learns to implement a task by only considering exam-
ples, without being programmed with any specific task rules.
ANN is based on a collection of nodes or units called neurons.
Neurons are the basic components of a neural network through
which communication takes place. A simple architecture of
ANN is depicted in Figure 3. A neuron receives input and
produces output based on its internal activation function [18].
The output of some neurons is the input to other neurons
forming a directed weighted graph. The functions as well as
the weights that compute the activation are altered by a process
known as learning. The parameters controlling the learning
for ANNs are the number of hidden layers, the learning rate,
and the maximum number of iterations. In hidden layers,
the number of neurons varies. ANN training model uses the
historical data to train itself and then make a prediction based
on new input data. Various activation functions are used in
ANN for computation, such as Softmax, Sigmoid, Rectified
linear unit, etc.
B. Deep Neural Networks
Deep neural networks (DNN) belongs to the ANN family.
They consist of multiple hidden layers between the input
and output layers [19], [20]. A comparison of DNN with
simple ANN is presented in Figure 3. The DNN processes
the input with mathematical manipulation to produce the
output, irrespective of whether the data relationship is linear
or nonlinear. The neural network is trained using a training
set, which results in the calculation of the probability of
each output. A complex DNN contains more layers than other
neural networks as shown in Figure 3.
C. Convolutional Neural Networks
Convolutional neural networks (CNN) belong to the family
of neural networks in deep learning. They are commonly used
for visual image processing, energy management in smart
grids, pattern recognition, etc. CNN is the updated version of a
Multi-Layer Perceptron (MLP). MLP is also known as a fully
connected layered network, which means every neuron in one
layer is fully connected to all neurons of the next layer. This
fully connected property often leads to an over-fitting problem
with MLP. However, CNN uses different approaches for data
regularization. Specifically, it exploits the hierarchical pattern
of data and assembles it to multiple simpler patterns. On the
basis of their translation invariance characteristics, CNN are
also known as shift invariant or space invariant ANNs [21].
CNNs are inspired by the biological process [22], in which
the neurons are fully attached to one another. It requires
little pre-processing compared to other image processing al-
gorithms. CNN works as a neural network, and it contains
an input layer, output layer, and hidden layers. The difference
between them is that CNN uses a series of different types of
hidden layers, i.e., convolutional layer, flatten layer, dropout
layer, pooling layer, fully connected layer, and normalization
layers. The input and output of the hidden layers are hidden
by an activation function. Rectifier linear unit (Relu) is mostly
used as an activation function in CNN. The activation function
sometimes involves the back-propagation method to produce

Fig. 3: Typical architectures of artificial neural networks (ANN) and deep neural networks (DNN)
a more accurate product or weight. The convolutional layer in
CNN is used to downscale the input data in such a way that
it becomes easy to process but the actual information remains
the same.
D. Recurrent Neural Networks
Recurrent neural networks (RNNs) are a special kind of
ANNs that are developed to process sequential data [23]. Usu-
ally, conventional networks provide training to each sample in-
dependently for each other; however, this type of independent
training is not sufficient, especially for data exhibiting tem-
poral relationships. RNN offers a solution to this problem by
taking inputs sequentially. Unlike other feed-forward ANNs,
they contain feedback connections in the units of the hidden
layer. Therefore, the RNN can perform temporal processing
and learn sequentially. Furthermore, unlike other ANNs, the
RNN uses a hidden layer as a memory for storing sequential
information. It is important to note that the RNN uses the
same parameters (like U, V, W shown in Figure 4) for each
layer instead of using different parameters for each layer like
traditional DNNs.
Figure 4 presents the RNN being unfolded into a full
network. For instance, if the RNN is used to make a prediction
based on the sequence of the last 6 data inputs, then the
network would be unfolded into a 6-layer neural network. In
RNN calculations, xt,ot, and stshow the input, output, and
hidden state at time t, respectively.
Fig. 4: An example of RNN architecture [24]
III. DEE P LEARNING IN ENERGY MANAGEMENT SYSTEMS
Smart microgrids nowadays integrates RERs, such as solar
panels and wind turbines, along with energy storage systems.
A typical architecture of a microgrid is depicted in Figure
2. Due to the intermittent power generation from renewable
resources, precise power forecasting has become crucial to
achieve efficient energy management. In this section, we will
uncover the deep learning based techniques used to predict
solar energy generation, summarized in Table I of this study.
A. Solar Energy Forecasting
Energy management in the residential or commercial sector
plays a vital role in enhancing grid stability and reliability.
When smart homes are integrated with RERs (e.g., wind tur-
bines, solar panels, etc.) for power generation, it is necessary to
predict the energy generation from these sources for efficient
energy management. Energy generation from these sources
may be predicted on the basis of 1 hour, 2 hours, 10 hours,
1 day etc. A lot of solar prediction studies are presented in
the literature and in this section, we have critically analyzed
them.
In [9], the authors have developed a solar forecasting
method using deep learning. In this work, 21 solar panels
are considered for electricity generation and one-day-ahead
prediction is performed by deep learning. Multi-layer percep-
tron (MLP) [25] is used as a base architecture in their work,
which consists of multiple layers as presented in Figure 5. In
addition, they have explored the use of deep belief networks
(a type of deep neural network) and two types of recurrent
neural networks. Finally, the deep learning forecasting results
are compared with each other as well as to physical models
showing promising results for recurrent neural networks.
Xwegnon et al. have proposed a statistical method for short-
term spatio-temporal prediction of solar energy generation
[10]. This paper considers the prediction for a very short time
horizon (1 to 6 hours). Distributed power plants are used in this
work as sensors and their spatio-temporal dependencies are
exploited to enhance the forecasting accuracy. Furthermore,
the computational complexity of the proposed model is low,

TABLE I: Summary of solar energy forecasting methods
Ref. Method(s) Compared method(s) Location Horizon Outcome/observation(s)
[9] Auto-LSTM MLP, ANN, DNN,
DBN, LSTM
Germany Hourly The newly proposed hybrid algorithm shows efficacy in terms
of accuracy; however, the DBN performence is close to the
proposed auto-LSTM forecasting method. [RMSE of proposed
method: 0.0713; compared method: 0.0714].
[10] Spatio-temporal
model
Auto-regressive, Ran-
dom Forest
France 15-minutes The proposed statistical model shows higher performance
in terms of accuracy and execution time over counterparts.
[nRMSE is improved up to 20% higher than the benchmark
methods].
[11] LSTM GRU, RNN, Naive DL France Hourly The LSTM forecasting model shows supremacy to compared
algorithms. [RMSE of proposed method: 0.2115 and compared
method: 0.2198]
[12] Gaussian process
regression-based
CNN
Persistence, ridge
regression, Fully-
connected NN
Oklahoma,
USA
Hourly The newly developed gaussian process regression based CNN
method shows efficacy in terms of minimum MAE. [MAE of
proposed method: 212642; compared method: 439952]
[14] DNN (‘SolarisNet’) ANN, SVR, Gaussian
Process Regression
India Hourly The proposed SolarisNet forecasting model presents higher ac-
curacy and it is validated through RMSE. [RMSE of proposed
method: 1.7661; compared method: 2.7930]
[15] Deep RNN FNN, SVR, LSTM Canada Hourly Results validate the performance of their proposed deep RNN
over counterparts in terms of RMSE. [Mean RMSE of pro-
posed method: 0.068; compared method: 0.18]
[32] ALHM model ANN, SVM - Hourly and
5-minutes
Based on GA and time-varying multiple linear model, the
proposed hybrid model is able to make precise prediction
of energy generated from solar panels [MAPE of proposed
method: 13.68; compared method: 20.39]
[33] Hybrid LSTM-RNN ANN, Bagged Regres-
sion Trees, Multiple
Linear Regression
Aswan and
Cairo, Egypt
Hourly The newly developed hybrid algorithm provides a very small
error rate compared to benchmarks. [RMSE of proposed
method: 82.15; compared method: 384.89]
[34] High-precision
Deep CNN
SVM, Random Forest,
Decision Tree, MLP,
LSTM
Tainan, Tai-
wan
Hourly The proposed high-precision deep CNN shows efficacy in
terms of minimum error rate. [Average MAE of proposed
method: 112.2640; compared method: 143.2721]
[35] DNN Ensemble SVR Oklahoma,
USA
Hourly The developed DNN uses minibatch training, dropout regu-
larization, and weight initialization to exploit and introduce
independent randomness in natural way. Results validate that
the DNN ensemble model is robust and has higher accuracy for
a single network. [Average MAE of proposed method: 209.09;
compared method: 222.52]
[36] Hybrid Gradient
Boosting Trees w/
feature engineering
technique
Quantile Regression
Forests
Porto, Portu-
gal
Hourly The presented work proposes a framework to extract features
from NWP grid using domain knowledge. In addition, they
proved that this information can enhance the forecast efficiency
in existing algorithms. [The proposed method shows average
point forecast improvement 16.09% over counterparts]
[37] SVM-RBF Linear Regression,
Past-predicts Future
Models
USA Hourly The SVM-RBF shows productiveness in terms of higher
accuracy. [The proposed method improves accuracy by 27%
over existing methods]
which makes it easy to use and appropriate for large scale
applications. Simulation results validate their proposed model
in terms of high accuracy as compared to other state-of-the-art
models. The authors of [11] have designed a prediction model
for solar irradiations, which is based on RNN. In particular,
they use two variants of RNN, namely gated recurrent unit
(GRU) and long short-term memory (LSTM) [26]. At the
end, simulations are performed to verify the performance of
RNN variants in terms of solar irradiation forecasting. Results
show that GRU and LSTM are more suitable for time series
prediction compared to simple RNN.
Another work proposes a solar forecasting model with nu-
merical weather prediction (NWP) and CNNs [12]. Moreover,
to train the CNN, a Gaussian process (sci-kit learn library
v0.19.0 [13]) is used to transform the incoming solar energy
values to the main grid. The proposed CNN is further able to
map the 6×6inputs to 31×31 output on the basis of the trans-
posed convolution operation. At the end, the proposed CNN
is validated through simulations and satisfactory accuracy is
presented as compared to three benchmark models: persistent
method, fully connected NN, and ridge regression methods.
Subhadip et al. have developed a deep NN, namely SolarisNet,
for global solar prediction [14]. They have used minimum
meteorological parameters, i.e., minimum temperature, maxi-
mum temperature, and hourly data of sunshine. Experiments
have been performed to check the adequacy of the proposed
SolarisNet model while data is taken from the meteorological
department of India. Experimental results validate that the
proposed model is more efficient as compared to existing
models, i.e., support vector regression (SVR) [27], Gaussian
process regression [28] and ANN [29], [30].
In [15], another solar energy forecasting method is devel-
oped using deep RNN (DRNN). The developed network is
trained, tested, and validated on real-time data obtained from
the National Resources of Canada [31]. Experimental results
are compared with other existing forecasting methods, which

|1
|2 H3
01H2
H1
Fig. 5: An example of Multi-Layer Perceptron (MLP)
show the effectiveness of their proposed scheme. The paper
[32] proposed a new adaptive learning hybrid model for short
and long term solar intensity prediction. To tackle the dynamic
and linear properties of the data, a time-varying multiple
linear model is developed. After that, a genetic algorithm
back-propagation neural network (GABPNN) is applied to
learn non-linear relationships in the data. Their proposed
novel adaptive learning hybrid model is able to capture the
temporal, linear, and non-linear relationships present in the
data. Simulation results validate that the proposed prediction
model outperforms multiple benchmark models in both long-
term and short-term forecasting of solar intensity.
A new solar power forecasting model is developed by
Abdel-Nasser et al. in [33], where the proposed model is based
on deep long short-term memory RNN (LSTM-RNN) and
considers the temporal changes during constructing prediction
models. They have analyzed the effectiveness of five different
LSTM models with different architectures. At the end, they
performed a comparative study to confirm the productiveness
of their proposed model. For comparison purposes, they con-
sidered some widely used forecasting models such as bagged
regression trees, NNs, and multiple linear regression. The work
[34] proposes a high precision deep CNN model, named So-
larNet, for solar irradiation forecasting. Real-time experiments
are performed to check the validity of the proposed model. It
is assured from results that the proposed SolarNet shows high
performance over counterparts in terms of higher accuracy.
Another work presented in [35] has developed two pre-
diction models for daily solar radiation and wind energy
predictions, which are based on DNNs. Input for these models
is taken from numerical weather systems. Kaggle data set is
used for model training and testing. This work also intro-
duces DNN ensembles to improve single DNN predictions
by lowering variance, while experiments validate that the
randomness in DNN training elements result in effective
and robust DNN ensembles. Another solar and wind energy
forecasting framework is presented in [36], where the authors
explore the information from a grid of numerical weather
predictions (NWP). Proposed methodology has considered
gradient boosting algorithm along with feature engineering
techniques, which extract the information from the NWP grid.
Furthermore, the authors have provided a comparison with
the methodology having only one NWP point for a specific
location. The results show that the forecasting accuracy is
improved, using MAPE, by 12.85% and 16.09% for wind
and solar energy, respectively. Another machine learning-based
solar energy prediction model is developed in [37], where the
authors have predicted weather information. They have also
performed a comparison with multiple regression methods to
show the effectiveness of their technique. Simulation results
show that their proposed machine learning-based method
achieves 27% higher accuracy than existing techniques.
IV. CURRENT CHALLENGES
Deep learning-based techniques have been considered a ben-
eficial means to enhance the performance of smart microgrids.
These are efficient methods that provide possible strategic
solutions for accurate forecasting of power generation from
solar panels. This section outlines some promising research
opportunities / directions of deep learning-based methods
implemented in solar energy prediction.
Big energy data analytics is an emerging, promising re-
search area, which needs special attention from the research
community. With the advent of sensor technology, wireless
communication, cloud computing, and advanced metering in-
frastructure, an enormous amount of data will become avail-
able, including energy consumption data, energy generation
data, temperature data, humidity data, etc. It is important to
make this data accessible to researchers and industry in order
to open new opportunities for improving the performance of
smart microgrids. The work [38] presents a comprehensive
vision for big data energy management, including energy
generation, transmission, distribution and transformation, and
demand-side management. They discuss in detail smart data
sources along with their characteristics. Furthermore, this
study has also proposed a process model for big data energy
management.
Machine learning and deep learning-based techniques de-
pend on historical data and perform future predictions on
the basis of old data. However, a heavy reliance on big data
requires a huge amount of storage devices. Furthermore, the
need for large-scale processing is another big challenge when
using deep learning-based methods. Unnecessary features and
data redundancy are the main causes of data complexity
and high computation. The computation time of redundant
training data is often much higher than training of clean data.
Machine learning techniques and other classifiers can be used
to remove the redundancy from data and make the training
data faster, while at the same time, help improve the accuracy
of classification and regression. So, a forecasting system is
needed that requires low processing along with low storage.
V. CONCLUSION
In this paper, we have presented a review on solar energy
prediction and its impact on smart microgrids. Energy gener-
ation from solar panels is intermittent in nature because of its
dependency on weather conditions, i.e., temperature, irradia-
tion, humidity, etc. For the efficient utilization of the generated
energy from solar panels, an accurate energy forecasting is

required. Machine learning and deep learning-based tools are
considered effective ways for energy forecasting on the basis
of historical data. So, we have presented a detailed review on
the current deep learning-based methods as well as reviewed
the efforts of researchers in enhancing the performance of
deep learning-based techniques in microgrids. Finally, we have
also unfolded the current challenges and future trends in
deep learning methods to enhance the forecast accuracy in
microgrids.
REFERENCES
[1] Mohanty, S., Patra, P. K., Sahoo, S. S., & Mohanty, A. (2017). Forecasting
of solar energy with application for a growing economy like India: Survey
and implication. Renewable and Sustainable Energy Reviews, 78, 539-
553.
[2] Aslam, S., Aslam, S., Herodotou, H., Mohsin, S., M., & Aurangzeb,
K. (2019). Towards Energy Efficiency and Power Trading Exploiting
Renewable Energy in Cloud Data Centers. In International Conference
on Advances in the Emerging Computing Technologies (AECT). IEEE.
[3] Khalid, A., Aslam, S., Aurangzeb, K., Haider, S., Ashraf, M., & Javaid, N.
(2018). An efficient energy management approach using fog-as-a-service
for sharing economy in a smart grid. Energies, 11(12), 3500.
[4] Woyte, A., Van Thong, V., Belmans, R., & Nijs, J. (2006). Voltage
fluctuations on distribution level introduced by photovoltaic systems.
IEEE Transactions on energy conversion, 21(1), 202-209.
[5] Enerdata, Global Energy Trends 2018. A step backward for the en-
ergy transition. https://www.enerdata.net/system/files/publications/global-
energy-trends-2018-edition-enerdata.pdf [Accessed on February 20, 2019]
[6] Aslam, S., Javaid, N., Asif, M., Iqbal, U., Iqbal, Z., & Sarwar, M. A.
(2018, June). A mixed integer linear programming based optimal home
energy management scheme considering grid-connected microgrids. In
2018 14th International Wireless Communications & Mobile Computing
Conference (IWCMC) (pp. 993-998). IEEE.
[7] Aurangzeb, K., Aslam, S., Herodotou, H., Alhussein, M., & Haider,
S. I. (2019, June). Towards Electricity Cost Alleviation by Integrating
RERs in a Smart Community: A Case Study. In 2019 23rd International
Conference Electronics (pp. 1-6). IEEE.
[8] Aslam, S. (2018). An optimal home energy management scheme con-
sidering grid connected microgrids with day-ahead weather forecasting
using artificial neural network. Masters Thesis, COMSATS University
Islamabad, Islamabad, Pakistan.
[9] Gensler, A., Henze, J., Sick, B., & Raabe, N. (2016, October). Deep
Learning for solar power forecasting—An approach using AutoEncoder
and LSTM Neural Networks. In 2016 IEEE international conference on
systems, man, and cybernetics (SMC) (pp. 2858-2865). IEEE.
[10] Agoua, X. G., Girard, R., & Kariniotakis, G. (2017). Short-term spatio-
temporal forecasting of photovoltaic power production. IEEE Transactions
on Sustainable Energy, 9(2), 538-546.
[11] Sorkun, M. C., Paoli, C., & Incel, ¨
O. D. (2017, November). Time
series forecasting on solar irradiation using deep learning. In 2017
10th International Conference on Electrical and Electronics Engineering
(ELECO) (pp. 151-155). IEEE.
[12] Chen, K., He, Z., Chen, K., Hu, J., & He, J. (2017, November). Solar
energy forecasting with numerical weather predictions on a grid and
convolutional networks. In 2017 IEEE Conference on Energy Internet
and Energy System Integration (EI2) (pp. 1-5). IEEE.
[13] Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B.,
Grisel, O., ... & Vanderplas, J. (2011). Scikit-learn: Machine learning in
Python. Journal of machine learning research, 12(Oct), 2825-2830.
[14] Dey, S., Pratiher, S., Banerjee, S., & Mukherjee, C. K. (2017). Solarisnet:
A deep regression network for solar radiation prediction. arXiv preprint
arXiv:1711.08413.
[15] Alzahrani, A., Shamsi, P., Ferdowsi, M., & Dagli, C. (2017, November).
Solar irradiance forecasting using deep recurrent neural networks. In
2017 IEEE 6th international conference on renewable energy research
and applications (ICRERA) (pp. 988-994). IEEE.
[16] Zia, M. F., Elbouchikhi, E., & Benbouzid, M. (2018). Microgrids
energy management systems: A critical review on methods, solutions,
and prospects. Applied energy, 222, 1033-1055.
[17] Hassoun, M. H. (1995). Fundamentals of artificial neural networks. MIT
press.
[18] Daniel, G. (2013). Principles of artificial neural networks (Vol. 7). World
Scientific.
[19] Bengio, Y. (2009). Learning deep architectures for AI. Foundations and
trends R
in Machine Learning, 2(1), 1-127.
[20] Schmidhuber, J. (2015). Deep learning in neural networks: An overview.
Neural networks, 61, 85-117.
[21] Zhang, W., Itoh, K., Tanida, J., & Ichioka, Y. (1990). Parallel distributed
processing model with local space-invariant interconnections and its
optical architecture. Applied optics, 29(32), 4790-4797.
[22] Matsugu, M., Mori, K., Mitari, Y., & Kaneda, Y. (2003). Subject
independent facial expression recognition with robust face detection using
a convolutional neural network. Neural Networks, 16(5-6), 555-559.
[23] Hopfield, J., J. (1984). Neurons with graded response have collective
computational properties like those of two-state neurons. Proceedings of
the national academy of sciences, 81(10), 3088-3092.
[24] Britz, D. (2015). Recurrent Neural Networks Tutorial. [online].
URL: http://www.wildml.com/2015/09/recurrentneural-networks-tutorial-
part-1-introduction-to-rnns/ [accessed June 28, 2019].
[25] Colchester, K., Hagras, H., Alghazzawi, D., & Aldabbagh, G. (2017).
A survey of artificial intelligence techniques employed for adaptive
educational systems within e-learning platforms. Journal of Artificial
Intelligence and Soft Computing Research, 7(1), 47-64.
[26] Bao, W., Yue, J., & Rao, Y. (2017). A deep learning framework for
financial time series using stacked autoencoders and long-short term
memory. PloS one, 12(7), e0180944.
[27] Chen, J. L., Liu, H. B., Wu, W., & Xie, D. T. (2011). Estimation of
monthly solar radiation from measured temperatures using support vector
machines–a case study. Renewable Energy, 36(1), 413-420.
[28] Salcedo-Sanz, S., Casanova-Mateo, C., Mu˜
noz-Mar´
ı, J., & Camps-Valls,
G. (2014). Prediction of daily global solar irradiation using temporal
Gaussian processes. IEEE Geoscience and Remote Sensing Letters,
11(11), 1936-1940.
[29] Amrouche, B., & Le Pivert, X. (2014). Artificial neural network based
daily local forecasting for global solar radiation. Applied energy, 130,
333-341.
[30] Yadav, A. K., & Chandel, S. S. (2014). Solar radiation prediction
using Artificial Neural Network techniques: A review. Renewable and
sustainable energy reviews, 33, 772-781.
[31] N. R. Canada. High-resolution solar radiation datasets. (2017). [Online].
Available: http://www.nrcan.gc.ca/energy/renewable-electricity/solar-
photovoltaic/18409
[32] Wang, Y., Shen, Y., Mao, S., Cao, G., & Nelms, R. M. (2018). Adaptive
learning hybrid model for solar intensity forecasting. IEEE Transactions
on Industrial Informatics, 14(4), 1635-1645.
[33] Abdel-Nasser, M., & Mahmoud, K. (2019). Accurate photovoltaic power
forecasting models using deep LSTM-RNN. Neural Computing and
Applications, 31(7), 2727-2740.
[34] Kuo, P. H., & Huang, C. J. (2018). A green energy application in energy
management systems by an artificial intelligence-based solar radiation
forecasting model. Energies, 11(4), 819.
[35] D´
ıaz–Vico, D., Torres–Barr ´
an, A., Omari, A., & Dorronsoro, J. R.
(2017). Deep neural networks for wind and solar energy prediction.
Neural Processing Letters, 46(3), 829-844.
[36] Andrade, J. R., & Bessa, R. J. (2017). Improving renewable energy fore-
casting with a grid of numerical weather predictions. IEEE Transactions
on Sustainable Energy, 8(4), 1571-1580.
[37] Sharma, N., Sharma, P., Irwin, D., & Shenoy, P. (2011, October). Pre-
dicting solar generation from weather forecasts using machine learning.
In 2011 IEEE international conference on smart grid communications
(SmartGridComm) (pp. 528-533). IEEE.
[38] Zhou, K., Fu, C., & Yang, S. (2016). Big data driven smart energy
management: From big data to big insights. Renewable and Sustainable
Energy Reviews, 56, 215-225.