International Journal of Modelling and Simulation

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tjms20

Forecasting of solar and wind power using

LSTM RNN for load frequency control in isolated
microgrid

Dhananjay Kumar , H. D. Mathur , S. Bhanot & Ramesh C. Bansal

To cite this article: Dhananjay Kumar , H. D. Mathur , S. Bhanot & Ramesh C. Bansal
(2020): Forecasting of solar and wind power using LSTM RNN for load frequency
control in isolated microgrid, International Journal of Modelling and Simulation, DOI:
10.1080/02286203.2020.1767840

To link to this article: https://doi.org/10.1080/02286203.2020.1767840

Published online: 30 Jun 2020.

Submit your article to this journal

Article views: 6

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tjms20
INTERNATIONAL JOURNAL OF MODELLING AND SIMULATION
https://doi.org/10.1080/02286203.2020.1767840

ARTICLE

Forecasting of solar and wind power using LSTM RNN for load frequency control
in isolated microgrid
a
Dhananjay Kumar , H. D. Mathura, S. Bhanota and Ramesh C. Bansalb
a
Department of Electrical & Electronics Engineering, BITS Pilani, Pilani, India; bDepartment of Electrical Engineering, University of Sharjah,
Sharjah, United Arab Emirates

ABSTRACT ARTICLE HISTORY

Renewable sources such as solar PV and wind are stochastic in nature, hence their integration with Received 16 October 2019
emerging isolated microgrid (MG) is challenging especially with regards to stability issues. An Accepted 8 May 2020
accurate prediction model of wind and solar sources is necessary to analyze the uncertainty in MG KEYWORDS
system and to encourage the reliable participation of wind and solar power in the energy market. Deep learning; time series
The advancement in deep learning methods has made it possible to develop a multi-step forecasting; LSTM recurrent
forecasting model unlike shallow neural networks (SNNs). The time series forecasting using SNN neural network; microgrid;
and Recurrent Neural Network (RNN) suffers from the problem of vanishing/exploding gradient load frequency control
while training. To eliminate this problem the long short-term memory (LSTM) RNN has been used
in this study for wind speed and solar irradiance prediction. The forecasted solar and wind power is
applied to analyze the load frequency behavior and the response of nonrenewable sources for
sudden rise and fall in load power demand and PI controller is used to mitigate frequency deviation
to ensure the stability of the MG power system.

1. Introduction varying continuously [4]. The electric load demand is

also not fixed and it also keeps on changing with time. In
The conventional power grid is undergoing through the
this scenario when three intermittent variables, i.e.,
significant reforms that will change its structure, control
solar, wind, and load power are continuously varying
techniques, and dynamic behavior. Looking to the cel-
in MG, the main challenge is to meet demand supply
lular structure of a large grid which is composed of
balance to maintain the frequency within safe limits.
many small-scale power system units and each of them
Inside the MG the optimum utilization of the renewable
have the ability to function independently during emer-
power increases the efficiency of a complete power sys-
gency conditions when there is no support from the
tem. Under such a dynamic scenario there is strong
central grid. These power system cells are termed as
need for a better technique that can predict the variation
‘Microgrids’ (MGs). A microgrid is defined as
in the intermittent variables such as solar, wind, and
a medium rated power system with local distributed
load power. The better prediction model for wind and
generations (DGs), storage systems, and loads with
solar power is very much useful in optimizing the
dedicated control arrangements to ensure the stability
energy harvested from RESs to reduce the fossil fuel
by continuously monitoring and regulating the voltage
consumption by secondary sources and it also helps in
and frequency [1,2]. To support the critical loads during
ancillary services and support [5]. In MG the secondary
emergency situation the capacity of generation installed
generation sources are DEGs (diesel engine generators),
in MG should exceed the peak demand [3]. The demand
micro turbines units (MTUs), fuel cell generators
for renewable energy sources (RESs)-based MGs are
(FCGs), and storage sources like battery, flywheel
increasing rapidly for remote areas, urban organiza-
energy storage systems (ESSs). The forecasting of
tions, and isolated communities in order to operate
renewable power may also contribute to analyze the
independently from the main grid and also to share
prefault stability of the MG by ensuring the proactive
their extra power to grid during peak demand.
participation of secondary sources. The forecasting pro-
Being the giant source of energy the wind and sun are
cess predicts the amount of power that may be harvested
available almost everywhere on the earth. Due to the
from RESs in a MG so that the back-up support from
stochastic nature of these RESs the power shared by
secondary power sources can be estimated and
primary sources (i.e. wind and solar) to MG keeps on
scheduled.

CONTACT Ramesh C. Bansal rcbansal@ieee.org Department of Electrical Engineering, University of Sharjah, Sharjah, United Arab Emirates
© 2020 Informa UK Limited, trading as Taylor & Francis Group
2 D. KUMAR ET AL.
With growing penetration of wind and solar power
to the main grid, transmission, and distribution com-
panies are experiencing instability concerns related to
the voltage and frequency and the power sharing
dynamics and its economic aspects are required to be
reanalyzed [6,7]. This concern is mainly due to inter-
mittent nature of renewable generation sources that is
dependent upon uncertain weather conditions. Better
prediction of wind and solar power will pave the way
for efficient and optimum utilization of renewable
energy sources [8]. The forecasted patterns of genera-
tion and load power for upcoming days are used to
formulate the policy for scheduling of generators and
planning the transactions in the electricity market to
balance the supply and demand of energy ensuring
stable operation of grid. The trends obtained after
forecast are utilized by transmission/distribution com-
panies, energy solution companies, power trade mar-
kets, and isolated power grids so that their power
transaction can be effectively scheduled, dispatched,
and monitored. Hence, there is a need for
a forecasting method with better accuracy and intelli-
gence and also forecasting based study of MG model
for stability analysis under varying load conditions.
The speed and direction of wind depend upon the
natural parameters like the Coriolis force, pressure-
gradient force, and frictional force to the Earth surface
[4]. The forecasted wind speed can be expressed in
terms of different wind components such as base
speed, gust speed, ramp speed, and noise [5]. The
work presented in [9] briefs about the three different
power forecast horizons, i.e., for an hour, for a day,
and day before. Several grid managing activities such
as unit commitment, day ahead markets, ramping
events, transmission scheduling, variability assessment,
planning, hedging, and asset optimization are related to
the forecast horizons. In [10] the different category of
time horizons has been defined which are considered in
forecasting, i.e., very short-term (up to 30 minutes),
short-term (few minutes to 6 hours), medium-term
(few hours to 24 hours), and long-term (more than
1 day). As per the regulation enacted by National
Energy Bureau (NEB) of China in 2011, for a wind
plant to plan the dispatch preparation the hourly pre-
diction of day-ahead forecasting is needed and also the
maximum error in daily forecast curve is limited to 25%
and the root mean square error (RMSE) of the all-day
forecast should be less than 20% [11]. The scheduling of
wind and solar power is at a nascent stage in India and
the forecast error margin is higher than ±30% [12]. For
variable speed wind turbines the forecasting of wind
speed has been presented in [13] for maximum power
point tracking and control.
Forecasting methods of wind and solar power can be
categorized as physical methods, statistical methods,
and machine learning (ML) methods [14]. Physical
methods are based upon the numerical weather predic-
tion (NWP) since it is related to the process of trans-
forming the wind and solar in to electric power [15].
Statistical methods applied for forecasting are time-
series-based approach, where predicted values can be
expressed as a linear or nonlinear function of the pre-
vious data [16,17]. The ML network models can map
a nonlinear and complex relationship between input
and output by applying various algorithms. Several
structures of artificial neural networks (ANNs) such as
feed-forward neural networks (FFNNs), support vector
machines (SVMs), and extreme learning machines
(ELMs) have been used by researchers to forecast wind
and solar power. However, deep learning has been used
rarely for multi-step sequential forecasting. The possible
reason behind it may be the large number of forecasting
steps that eventually reduces the accuracy. The classifi-
cation and sub classification of machine learning tech-
niques for various applications is given in [18]. The
conventional shallow neural network (SNN) models
suffer with slower convergence and overfitting problem
and they can be easily trapped in to the local minimum
[19]. To solve this issue, researchers have applied the
deep neural network (DNN), which has many variants
like the deep belief network (DBN), convolutional
neural network (CNN), and recurrent neural network
(RNN) [20,21]. RNNs are more powerful as compared
to ANNs in terms of feature extraction and finding the
hidden and high level invariant structures from data.
The output of RNN depends upon the present input and
previous step observation and sometimes while learning
it faces the difficulty of long-term dependency. This is
referred as vanishing/exploding gradient problem [22].
To overcome this problem, an advance RNN has been
proposed which is termed as long short-term memory
(LSTM) [23,24]. The LSTM network gives better result
for recursive multistep forecasting by eliminating the
vanishing gradient and gradient explosion problem. It
completely utilizes the large amount of training data for
forecasting, clustering and classification.
The variational mode decomposition (VMD) and sin-
gular spectrum analysis (SSA) techniques have been used
in [25] for recursive small multistep wind speed forecast-
ing. The work reported in [26,27] predicts the upcoming
workload on the server using ML approaches like self-
organizing feature map (SOFM), support vector machines
(SVMs), k Nearest Neighbors (kNN) neural networks, and
nature inspired algorithms. In [28] two-tier architecture
using kNN has been presented for financial time series
prediction. Forecasting of mean hourly wind speed data

using time series analysis has been presented in [29] where
the forecasting have been accomplished using linear
ARIMA models and feed forward artificial neural net-
works. These models were tested on two different datasets
and the ANN model outperformed the ARIMA model.
A local recurrent neural network (RNN) model trained
using Recursive Prediction Error (RPE) algorithm for fore-
casting of wind power and speed been presented in [30]. In
[31] review of forecasting methods for solar irradiance has
been summarized. Several models were introduced
for hour ahead forecast in [32], where ANN-based model
was optimized by genetic algorithm. To estimate the solar
radiation by ANNs, review of several techniques has been
presented in [33] that discuss the nature of the ANNs and
the comparison of different models for solar radiation
estimation. An approach for forecasting power generation
24 hours ahead using a radial basis function network
(RBFN) is proposed in [34] for online training of experi-
mental roof top PV system. A hybrid iterative method for
short-term wind speed forecast is presented in [35]. Based
on variational mode decomposition method a multi-scale
wind speed forecasting is presented in [36].
To ensure the stability of MG power system in
response to the disturbance such as unbalancing
between demand and supply the controller play a very
significant role by regulating the frequency within a safe
operating range. In [37] a neural network based upon
the Gaussian mixture model has been proposed to fore-
cast the short-term wind power generation. Using fuzzy
PI controllers in [38] the FC and AE hybrid system are
being controlled dynamically to achieve active power
balancing. A control system design has been presented
in [39] for under frequency load shedding in microgrid
utilizing the power generated from renewable energy
resources. Various optimization methods for load fre-
quency control in autonomous MG is presented such as
grasshopper optimization algorithm for fractional-order
PI-D controller [40], reinforcement learning-based con-
trol [41], and invasive weed optimization for automatic
generation control [42]. The work in [43] presents an
intelligent tuning method of fuzzy logic controller
(FLC) using particle swarm optimization (PSO). The
frequency control in an AC microgrid using online
PSO-based fuzzy tuning approach has been presented
in [44]. For energy management in smart grids a firebase
cloud messaging (FCM) system has been presented in
[45]. In [46], literature survey has been presented
describing different intelligent forecasting techniques
for load demand in power system. Design of two degree
of freedom-internal model control [47] and perfor-
mance analysis of a cascade controller [48] has been
presented for load-frequency control purpose.
INTERNATIONAL JOURNAL OF MODELLING AND SIMULATION 3
The survey of literature in above paragraph clearly indi-
cates the strength of deep learning (DL) for sequential time
series forecasting. Since, the RNN suffers with the difficulty
of long term dependency, i.e., vanishing/exploding gradi-
ent problem, hence to solve this problem an advanced
RNN, i.e., long short-term memory recurrent neural net-
work (LSTM-RNN) has been applied in this work to fore-
cast the solar irradiance and wind speed. The predicted
solar radiation and wind speed have been utilized to calcu-
late the forecasted power from renewable energy sources.
In the survey of literatures it has been also found that the
forecast-based load frequency controller analysis is rarely
investigated. Forecasted renewable power based load fre-
quency study is an effective way of studying the stability in
MGs. The forecasted power is injected to small signal
model of microgrid (MG) for load frequency control
study. The varying load demand in MG results into fre-
quency excursion from nominal value that may destabilize
the MG power system. This paper introduces a novel
approach of renewable power prediction based load fre-
quency control analysis to study the stability of MG in the
presence of intermittent renewable energy sources and
dynamic load demand. For forecasting the power from
renewable energy sources the hourly data of solar radiation
and wind speed have been taken from [49] for training the
LSTM-RNN. Section 2 discusses the architecture of LSTM-
RNN utilized for training the network with sequential data
of wind speed and solar irradiance. The section 3 presents
the different steps involved in training the LSTM-RNN
network with hourly data of solar irradiance and wind
speed. The real data is obtained from a Solar Park located
in Charanka village of Patan district under Gujarat Energy
Transmission Corporation Limited. Section 4 describes the
MG model for small signal analysis that integrates the
power obtained from solar and wind sources, auxiliary
sources like DEG, MT, FCG, BESS, and FESS. The simula-
tion results of load frequency control for regulating the
deviation in frequency of MG caused due to the change
in power of load and renewable generation (wind and
solar) is presented in section 4. The section 5 briefs the
conclusion drawn from simulation results.
2. LSTM RNN-based forecasting
The property to learn the long-term dependency by
RNNs makes it the more powerful network as compared
to other neural networks that have been applied for time
series forecasting. However it suffers with vanishing/
exploding gradient. In deep neural network the hidden
layers and corresponding time steps are related to each
other through multiplication so the derivatives/gradi-
ents/errors are more sensitive toward vanishing or
4 D. KUMAR ET AL.
Figure 1. Generalized structure of a microgrid (MG).
exploding. In order to solve this issue, the RNNs have
been modified to its extended version that is termed as
LSTM network. It was proposed by the German
researchers Sepp Hochreiter and Juergen Schmidhuber
as a solution to the vanishing gradient problem in
1997. The LSTM-RNN is very useful where it is impor-
tant to learn the past behavior and see its effect on
improving the accuracy in prediction of the future
values. The LSTM-RNNs are categorized on the basis
of its application. The regression output network is used
to forecast the single step or multi step time series data.
In this work sequence-to-sequence regression network
of LSTM RNN has been applied for forecasting the solar
and wind power for integration with the renewable
sources to MG. Figure 1 shows a simplified structure
of MG power system that mainly consists of power
sources, electrical load, and controller. Later the load
frequency control study has been done using forecasted
power applying step variations in load demand to study
the stability of MG power system.
2.1. Sequence-to-sequence LSTM RNN
The LSTM-RNN is trained in such a way that for every
iteration in input time step it learns to predict the value of
upcoming time step. The outcomes of the sequence-to-
sequence LSTM-RNN method are the training sequences
with values shifted by one time step. Figure 2 shows the
architecture of LSTM network for sequence-to-sequence
regression. The first layer inputs the sequence data with
specified size and number of features to the LSTM net-
work layer.
Figure 2. LSTM network architecture for sequence-to-sequence
regression output.
Figure 3. Long short-term memory block architecture [50].
2.2. LSTM network Layer
For implementing Deep Learning the LSTM networks are
utilized. LSTM networks are advanced RNNs with addi-
tional feature of preserving the information from the last
step. Unlike simple RNN, LSTM cells contain memory
for every time step. The LSTM block architecture is
shown in Figure 3. An LSTM block typically has
a memory cell, input gate, output gate, and a forget gate
in addition to the hidden state in traditional RNNs. An
LSTM layer learns long term dependencies between time
steps in time series and sequential data. The weights and
biases to the input gate control the extent to which a new
value flows into the cell. Similarly, the weights and biases
to the forget gate and output gate control the extent to
which a value remains in the cell and the extent to which
the value in the cell is used to compute the output activa-
tion of the LSTM block, respectively.
3. Training the LSTM network
The LSTM-RNN is trained by input data sequence. For
predicting the values of future sequence, the function is
trained to predict for one time step at a time and then
updating the network state after every prediction. The
complete data set is partitioned as training and test data.
The network is trained with 90% of the complete data
sets starting from first data and tested with 10% data
present at the last. For better fitting and to protect the
training from diverging, training data sequence is stan-
dardized to have zero mean and unity variance. The
pseudo code for the training the LSTM for time series
forecasting is shown as follows [50].
1: Loading the input data sequence
1(a): Load the sequence data set
1(b): partition the training data (Dtrain ) and test
data (Dtest )
2: Standardize the training data set
2(a): find mean (μ) and variance (σ) for training
data (Dtrain ).
2(b): get standardized dataset by (1), i.e., Dstd ¼
ðDtrain  μÞ=σ
2(c): specify the predictors (Xtrain ) and responses (Ytrain ).
 INTERNATIONAL JOURNAL OF MODELLING AND SIMULATION 5

The predictors are training sequences without the

final time step given by Equation (2) and responses are
the training sequences with values shifted by one time
step and is can be calculated by Equation (3).
3: Define LSTM network architecture with options
specified in Table 1.
4: for n epochs and ith iteration do
train the network (L) with the specified training
options
end for
5. Forecast future time steps
5(a): Standardize the test data using the same para-
meters as the training data using Equations (4) and (5). Figure 4. Training progress of LSTM network for solar irradiance
5(b): for i = 2: N (N = number of time steps in test and corresponding loss function.
data)
predict and update state
end
5(c): using Equation (6) find the unstandardized ser-
ies of the predictions (Ypred ) using the parameters from
step 2(a) and standardized prediction Ypred .
6. Calculate the validation and loss function.
Following the above pseudocode the predicted values
have been obtained. If the actual values of time steps
between predictions are known, then one can update the
network state with the observed values instead of the pre-
dicted values. This will yield more accurate prediction since
the network state is updated with the observed values
instead of the predicted values. To implement the forecast- Figure 5. Training progress of LSTM network for wind speed and
ing of solar irradiance and wind speed the MATLAB plat- corresponding loss function.
form have been used for this study. An LSTM regression
network has been modelled by defining the LSTM-RNN
layer with training options presented in Table 1. The train-
ðDtrain  μÞ
ing progress report is shown in Figure 4 for solar irradiance Dstd ¼ (1)
σ
hourly data and in Figure 5 for wind speed hourly data of
January month (1st to 31st). Total number of data points are Xtrain ¼ Dstd ð1 : end  1Þ (2)
744 (31 days *24 hours, i.e., 744 data points denoting hourly
wind speed and solar irradiance values). Ytrain ¼ Dstd ð2 : endÞ (3)
Equations (1)–(6) have been used for data processing at
different stages of forecasting. They are mainly used for ðDtest  μÞ
data set preparation, training the network, prediction, and Dteststd ¼ (4)
σ
testing.
Xtest ¼ Dteststd ð1 : end  1Þ (5)

Table 1. Training options for forecasting by LSTM model [50].

Ypred ¼ σ  Ypred þ μ (6)
Training option Value
Solver ‘adam’ (Adaptive moment estimation)
Hidden layer 200
Activation function SOFTMAX 3.1. Forecasting future time steps
Maximum Epoch 250
Gradient Threshold 1
Initial Learning Rate 0.005
The future values of multiple time steps are obtained
Learn rate schedule Piecewise using the MATLAB function predictAndUpdateState.
Learn rate drop period 125 This works as predicting the values at particular time
Learn rate drop factor 0.2
Iteration per epoch 1 step, one at a time and then the network state is updated
for every prediction. Thus the past predictions are
6 D. KUMAR ET AL.
Figure 6. Observed and forecasted pattern for solar irradiance
time series data for 31 days.
Figure 7. Observed and forecasted pattern for wind speed time
series data for 31 days.
utilized as an input to the function for each prediction
step. Firstly, standardize the test data to have zero mean
and unity variance. To initialize the network state, first
predict on the training data. Next, make the first pre-
diction using the last time step of the training response
and loop over the remaining predictions and input the
previous prediction to predictAndUpdateState.
Afterward, de-standardize the prediction using the
mean and variance calculated earlier. The RMSE is
calculated and the training data with the forecasted
ones are plotted in Figures 6 and 7 for solar and wind
time series respectively.
The RMSE calculated after training for solar irradi-
ance data is 11.62 and for wind speed data is 1.15. The
higher value of RMSE for solar is due to the fact that for
half of the day time the solar has zero output while the
wind is having nonzero output for the complete day
(24 hours). The observed values of solar irradiance and
wind speed is compared with forecasted values, i.e., the
test (observed) data which is 10 % of the complete dataset
(last 3 days or 72 hours out of 31 days or 744 hours of
January month) is plotted with the forecasted values as
shown in Figures 8 and 9 with RMSE values.
3.2. Updating the network state
While training the LSTM-RNN the predicted values
of previous step is feedback to the hidden layer. The
Figure 8. Forecasted and observed values (test data) for solar
irradiance and RMSE (3 days).
Figure 9. Forecasted and observed values (test data) for wind
speed and RMSE (3 days).
network state is then updated with current step
input and predicted value of previous step. As we
have access of the observed data hence to improve
the training accuracy, instead of updating the net-
work with predicted values the observed values are
feed to the network. To execute this step the net-
work is reset first and then the network state is
initialized by predicting on the training data. The
advantage of network state reset is that it does not
allow the past prediction to affect the predictions
done on the new set of data. The prediction is
destandardized using mean and standard deviation
calculated earlier and then RMSE is recalculated.
The updated forecast for solar irradiance and wind
speed is shown in Figures 10 and 11 with reduced
RMSE values of 4.45 and 0.17 for solar irradiance
and wind speed respectively.
To analyze the improvement in the accuracy of fore-
casting, a comparison is made between RNN LSTM
based forecasting and other techniques applied earlier.
The comparison parameter is the RMSE. Table 2 sum-
marizes the RMSEs obtained while forecasting the wind
speed and solar irradiance using proposed and existing
techniques.
 INTERNATIONAL JOURNAL OF MODELLING AND SIMULATION 7

sources and storage sources are controlled to maintain

Figure 10. Forecast after updating the network with observed
data and RMSE for solar irradiance.
the demand supply balance by regulating the MG fre-
quency within the allowable limits. This will ensure the
stability of MG power system along with preventing the
under frequency or over frequency load shedding pro-
blem. Wind and solar power are controlled by nature
variables and it hardly participates in the frequency
control operation of MG. Any mismatch in the total
generated power and load demand causes the fluctua-
tion in MG frequency which will degrade the power
quality and may lead to failure of whole MG power
system. Hence, the continuous monitoring of frequency
deviation must be done and the controller should take
the suitable action to mitigate the frequency deviation
within the specified time whenever there is sudden
change in load demand or generation. In this paper
a mathematical model of MG has been taken assuming
suitable inertia coefficient and load damping constant.
All the generation sources are taken by assuming a first-
order transfer function with appropriate time constant.
The net power deviation is fed to MG model and the
corresponding frequency deviation signal is monitored.
The frequency deviation is feedback to the controller
which actuates the secondary sources. In response to
control signal the secondary sources changes their
power share to microgrid so that the net power differ-
ence is made to zero and thus frequency deviation can
be brought within the allowable range.

Figure 11. Forecast after updating network with observed data

and RMSE for wind speed. 4.1. Power from solar and wind in islanded MG
The Equation (7) shows relation for calculating the power
Table 2. RMSE comparison for LSTM-RNN and other forecasting output of specified PV system (PPV ) whose parameters
methods.
are taken from [56–58]. In this case the conversion effi-
Forecasting parameter Method (reference) RMSE
ciency has been taken as 10%, ambient temperature
Solar irradiance ANN [51] 8.24 
ARMA [51] 8.54 is25 C. The base power is 1 MW, hence the PV power
LSTM-RNN (Proposed) 4.45 in per unit system is calculated after dividing the fore-
SARIMA (Computed) 8.73
ARIMA [52] 33.39 casted solar power by base power. Figure 12 shows the
P-ELM [53] 42.62 forecasted solar PV power in per unit. Similar to solar PV
Wind speed ANN (Computed) 4.32
LSTM-RNN (Proposed) 0.17 power obtained from the forecasted values of solar radia-
SARIMA [54] 0.21 tions in the above paragraph, the per unit power obtained
ARIMA [55] 0.34
NARX model [55] 0.39
from the wind turbine generator can be found by putting
Hybrid (ARIMA+ANN) [56] 0.49 the values of wind speed in Equation (8). The mechanical
power (Pm ) obtained from Wind Turbine Generator
(WTG) is expressed in Equation (9) and shown in
4. The microgrid model Figure 13.
The primary sources of power in MG are solar and wind PPV ¼ ηSφf1  0:005ðTa þ 25Þg (7)
power. Once the forecasted wind speed is available the
wind profile is given to the wind turbine model to ρA
Pm ¼ Cp ðλ; βÞ Vwind 3 (8)
convert it into electrical power. Similarly, the solar irra- 2
diance profile is converted to electric power assuming where,
suitable conversion parameters. In MG, the secondary Pm = mechanical power output of wind turbine,
8 D. KUMAR ET AL.

and FC can be considered as an actuator in MG model.

The power injection from diesel generator and fuel cell
is regulated by control signal from the PI controller to
restore the frequency deviation within permissible lim-
its. The total power generation in MG comprises the
power output from primary sources such as wind and
solar PV, power from the secondary sources, i.e., fuel
cell and diesel generator, and power from storage
sources like FESS and BESS. The gross power generated
is expressed by Equation (10). A fraction of power from
Figure 12. Power obtained using forecasted solar radiation. solar and wind is sent to AE which produces hydrogen
for FCG whenever there is surplus power is available
[59]. The electrolyzer breaks the water into hydrogen by
absorbing the power oscillations in WTG and PV. Since,
electrolyzer is consuming power that is represented by
assuming a negative sign. The power from storage
sources has dual nature, i.e., during normal case when
there is a balance between demand and supply they are
taking power from MG for charging and as soon as there
is sudden rise in load demand then they get discharged,
i.e., deliver power to grid for shorter duration and again
return back to charging state. Hence most of the times
Figure 13. Power obtained using forecasted wind speed. they act as a load to grid, hence the negative sign is there
in equation. Sometimes, when they contribute power to
A = turbine swept area (m2 ) MG so there will be positive sign in the equation.
Vwind = wind speed (m=s),
PTotal ¼ PWTG þ PPV  PAE þ PFC þ PDEG  PBESS
Cp = performance coefficient of wind turbine
 PFESS
ρ = air density (kg=m3 ), λ = tip speed
ratio (bladetipspeed=windspeedÞ (10)
β = blade pitch angle (degree) Equation (11) represents that the net power changes in
The normalized expression of Equation (8) is given sources must be zero to ensure the stability of the MG
by Equation (9). power system.
ρA
Pm pu ¼ Kp Cp pu ðVwind pu Þ
3
(9) ΔPTotal þ ΔPWTG þ ΔPPV þ ΔPAE þ ΔPFC þ ΔPDEG
2
þ ΔPBESS þ ΔPFESS ¼ 0
where,
(11)
Pm pu = per unit (p.u.) value of nominal power for
particular values of ρ and A. The input to the MG power system model is the power
Cp pu = performance coefficient in per unit of the difference (ΔPerror ) between change in demand
maximum value ofCp . (ΔPdemand ) and total power generated (ΔPTotal ). The
Vwind pu = wind speed in per unit of base wind speed. Equation (12) is used to calculate the power mismatch
Kp = power gain for Cp = 1 p.u. and Vwind pu = 1 p.u., subjected to MG.
since Kp should be less than equal to 1 so for this case it
ΔPerror ¼ ΔPdemand  ΔPTotal (12)
has been assumed to be 0.9.
A hybrid power system can be modelled by equivalent
inertia coefficient ‘M’ and load damping constant ‘D’
4.2. Power and frequency deviation in MG
with an inherent time delay between frequency and
Apart from uncontrolled RESs the other sources that power deviation [60]. Here the delay is neglected to
can be fully controlled are diesel engine generator assume a linear approximation. The input (i.e., power)
(DEG), fuel cells (FCs), and aqua electrolyzer (AE). error and output (i.e., change in frequency) can be
The driving fuel for FCs are hydrogen and it is gener- expressed in a transfer function form, as is given by
ated by AE. The oscillations in wind power are absorbed Equation (13). It correlates the change in power to
by AE to produce hydrogen needed by FCs. The DEG frequency deviation of MG system.

Δf 1 Table 3. Parameters of MG model.
Gs ðsÞ ¼ ¼ (13)
ΔPerror D þ Ms
4.3. Load frequency control in MG
To study the small signal variation in frequency and its
effect on the controlled sources the first-order transfer
function model has been assumed for modelling all the
power sources as shown in Figure 14. The wind turbine
generator is fed with the predicted wind profile and it
gives the electrical power coming out of WTG. Similarly,
solar PV model is fed with predicted irradiance profile
and it yields the solar power. The power coming from
primary sources are collectively fed to MG. The difference
in power of load and source is fed to inertia model of MG
power system to yield the net frequency excursion. The
frequency can be considered as the indicator of the bal-
ance between generation and load power. Hence, the
frequency deviation must be strictly regulated by
a controller for ensuring the stability. The parameters
used for making the closed loop transfer function model
of MG are summarized in Table 3 with their typical
values [61].
The frequency control in power system is further
categorized into three parts, i.e., primary, secondary,
Figure 14. The mathematical model of microgrid for load fre-
quency control study.
INTERNATIONAL JOURNAL OF MODELLING AND SIMULATION 9
Parameter Value Parameter Value
D (p.u/Hz) 0.012 TDEG (s) 2
M (p.u-sec) 0.2 TFC (s) 4
TBESS (s) 0.1 TFESS (s) 0.1
TPV (s) 1.8 TWTG (s) 1.5
TAE (s) 0.5 KP, KI 1.5, 2.5
and tertiary control. The primary control is directly
related to system inertia and it is done inherently by the
distributed energy resources (DERs) after the initiation of
disturbance in MG. The primary control action is taken
quickly within few second after the disturbance to limit
the high rate of change of frequency. The secondary
control is applied after primary to restore the frequency
by regulating the power from auxiliary sources. For test-
ing the controller response a step disturbance is applied to
MG for analyzing the load frequency behavior and other
controlled sources responses.
4.2. Results for secondary load frequency control in
MG
Forecasted values of solar irradiance and wind speed
obtained in section 3 are applied to the small signal
model of MG power system. The relationship of fre-
quency deviation with respect to change in power is
given by the transfer function Gsys ðsÞin section 4.1. The
MG system is subjected to step load disturbances and the
corresponding frequency deviation has been analyzed.
The effect of load rise and fall on the MG frequency has
been plotted along with the behavior of controlled
sources. The results related to step rise in load which
has been applied at 6th hour of the day are explained as
scenario 1. The results related to step rise in load applied
at 20th hour of the day are explained as scenario 2.
Scenario 1: A step rise in load at t = 21,600 seconds
(6:00:00 Hrs.) is applied to the MG power system
The MG power system is subjected to load power
which is also an intermittent parameter. The electric
load demand keeps changing during the day based
upon the need of the consumer. The load demand
pattern and the renewable energy source generation
are stochastic in nature and it is required that this
imbalance should be within ± 0.005 per unit Hz to
ensure the power system stability. To see the controller
response a step rise in load has been applied at 6th hour
of the day. The controller takes frequency deviation as
an input and controls three artificial sources, i.e., micro-
turbine, fuel cell and diesel engine generator. These
sources are modelled with their equivalent transfer
function model. The simulation results for scenario 1
have been plotted in Figure 15.
10 D. KUMAR ET AL.
Figure 15. Scenario 1 when a step load rise of 0.2 p.u is applied
at t = 21,600 seconds (a) step load rise (p.u.), (b) frequency
deviation (p.u. Hz), (c) micro turbine power (p.u.), (d) fuel cell
power (p.u.), (e) diesel engine generator power (p.u.), and (f)
BESS/FESS power output (p.u.).
When there is rise in load demand as shown in Figure
15(a) the frequency falls out of nominal range as shown
in Figure 15(b). To mitigate the frequency deviation the
controller responds and commands the artificial sources
to increase their power share to MG as shown in Figure
15(c-e) and during this period the storage sources
(BESS/FESS) are discharged as shown in 14(f).
Scenario 2: A step fall in load at t = 72,000 seconds
(20:00:00 Hrs.) is applied to the MG
A step fall in load has been applied at 20th hour of
the day. The simulation results for scenario 2 in Figure
16 show that when there is fall in load demand of 0.2 p.
u. as shown in Figure 16(a) the frequency shoots up
from its nominal value as shown in Figure 16(b). To
mitigate the frequency deviation the controller responds
and sends control signals to the artificial sources to
reduce their power share as shown in Figure 16(c-e),
and during this period the storage sources (BESS/FESS)
are charged as shown in Figure 16(f). This leads to
restoration of frequency in MG power system.
5. Conclusion
The implementation of LSTM-RNN based deep learning
method has been applied for multistep sequence to
Figure 16. Scenario 2 when a step load fall of 0.2 p.u. is applied
at t = 72,000 seconds (20th hour) (a) step load fall (p.u.), (b)
frequency deviation (p.u. Hz), (c) microturbine power (p.u.), (d)
fuel cell power (p.u.), (e) diesel engine generator power (p.u.),
and (f) BESS/FESS power output (p.u.).
sequence forecasting of solar and wind power. This
method is based on training, predicting and updating
the RNN states using memory cells in LSTM network.
The LSTM based forecasting is used to eliminate the
vanishing/exploding gradient problems that may occur
with RNN based forecasting. The RMSE obtained with
LSTM based forecasting of wind speed is found least
among the other reviewed techniques while for solar
irradiance forecast the RMSE obtained with RNN-LSTM
is also found very less. Also, the feature of updating the
LSTM-RNN with observed values instead of predicted
values further reduces the RMSE in forecast process to
yield better prediction. The RMSE for solar forecasting is
high in comparison to wind speed forecasting since the
mean and the standard deviation of the solar irradiance
data is high for solar. A more accurate forecasting for
wind speed and solar radiation is obtained with LSTM-
RNN. The day ahead forecasted wind and solar power is
fed to mathematical model of MG and load frequency
behavior has been studied. The close loop behavior of
auxiliary sources in MG is also studied to achieve the
demand supply balance after a step load disturbance of
20% in MG. A proportional and integral (PI) controller is
applied for actuating the secondary sources to restore the
frequency deviations in MG caused due to sudden change

in power demand. The forecasting error (RMSE) of wind
speed and solar irradiance is well within the acceptable
range. The proposed forecasting based MG model can be
utilized for analyzing the wind and solar integration to
MG and load frequency control study.
Disclosure statement
No potential conflict of interest was reported by the authors.
Funding
This research work is supported by the Department of Science
and Technology, New Delhi under the ICPS scheme through
letter no. [DST/CPS/CLUSTER/IoT/2018/General].
Notes on contributors
Dhananjay Kumar received the B.E. degree from Rajiv
Gandhi Proudyogiki Vishwavidyalay (RGPV), Bhopal, India
in 2009 and M.E. (Power Electronics) degree from Samrat
Ashok Technological Institute, Vidisha under RGPV Bhopal
in 2013. He is currently a full-time research scholar and
pursuing Ph.D. degree at department of Electrical and
Electronics Engineering, Birla Institute of Technology and
Science, Pilani, Pilani campus, Rajasthan, India. His research
interests include small signal modelling and stability control
in microgrids, renewable energy integration to MG, forecast-
ing of renewable power, robust control synthesis for fre-
quency control in microgrids. He has authored 3 research
articles in peer-reviewed international journals of repute and
his 7 papers have been presented in IEEE sponsored interna-
tional and top national conferences.
H. D. Mathur received B.E. degree from Nagpur University,
Nagpur, India, in 1998; M.E. degree from Malaviya Regional
Engineering College, Jaipur, India, in 2000; and the Ph. D.
degree from Birla Institute of Technology and Science (BITS),
Pilani, India in 2007 and He was Post-Doctoral Fellow in
Supelec, Paris, France in 2013. He was also visiting scientist
to Centralesupelec, France in May- June of 2015 and May-
June of 2019. Currently, He is Associate Professor in the
Department of Electrical and Electronics Engineering, Birla
Institute of Technology and Science, Pilani. He is having
teaching and research experience of more than 20 years.
Prof. Mathur is Associate Editor of IET Renewable Power
Generation as well as in editorial board and reviewer of
various journals mainly in field of electrical power system,
renewable energy. He is a Senior member IEEE, Fellow of
Institution of Engineers (India), Chartered Engineer and life
member of Indian Society of Technical Education. He has
published more than 65 papers in national/international jour-
nals and conference proceedings. His research interests
include power system control of isolated and interconnected
power systems, power system optimization, automatic gen-
eration control and artificial intelligence techniques applica-
tions in power systems and distributed generation (DG) with
grid interconnection issues. He is handling research projects
related to area of electrical power system, renewable energy
INTERNATIONAL JOURNAL OF MODELLING AND SIMULATION 11
integration, energy management in V2G etc. funded by Govt.
agencies.
S. Bhanot obtained B.E. (Hons) Mechanical Engineering and
M.Phil (instrumentation) from BITS, Pilani, and PhD from
IIT Roorkee (then University of Roorkee). She has 17 years of
teaching experience in BITS Pilani, and 19 years in Thapar
University. She is presently Professor at EEE Department
BITS Pilani. Her teaching and research areas are sensors,
industrial instrumentation & automation, biomedical instru-
mentation, application of AI techniques in process modelling,
control, image processing. She has been guiding projects/
theses at first degree, higher degree and PhD level and has
published research articles in international, national journals
and conferences. She has published a book “Process Control,
Principles and applications” with Oxford University press.
She is reviewer of many books, research papers and sponsored
projects. She received best teacher award at Thapar university
and nominated by globally dispersed BITS alumni as a teacher
that they would especially like to recognise.
Ramesh C. Bansal has more than 25 years of teaching,
research, academic leadership, and industrial experience.
Currently he is Professor in the Department of Electrical
Engineering at University of Sharjah, Sharjah, United Arab
Emirates. In previous postings, he was Professor and Group
head (Power) in the Department of Electrical, Electronic and
Computer Engineering at the University of Pretoria, South
Africa and worked with the University of Queensland,
Australia; University of the South Pacific, Fiji; Birla Institute
of Technology and Science, Pilani, India; and Civil
Construction Wing, All India Radio. Prof. Bansal has pub-
lished over 300 journal articles, presented papers at confer-
ences, and has published several books and chapters in books.
He has Google citations of over 9000 and h-index of 44. He
has supervised 20 PhD and 4 Post Docs and currently super-
vising several PhD students. He is an Editor/Associate Editor
of member many reputed journals including IEEE Systems
Journal, IET-Renewable Power Generation, Technology and
Economics of Smart Grids and Sustainable Energy. He is a
Fellow, and CP Engg. IET-UK, Fellow Institution of Engineers
(India), and Senior Member IEEE. He has diversified research
interests in the areas of Renewable Energy, Power Systems and
Smart Grid.
ORCID
Dhananjay Kumar http://orcid.org/0000-0002-5793-1418
References
[1] Jose J, Kowli A, Bhalekar V, et al. Dynamic
microgrid-based operations: A new operational para-
digm for distribution networks. IEEE Innov Smart Grid
Technol. 2016 Nov;28:564–569.
[2] Bansal RC, Editor. Handbook of distributed generation:
electric power technologies, economics and environ-
mental impacts. Cham, Switzerland: Springer; 2017.
[3] Parhizi S, Lotfi H, Khodaei A, et al. State of the art in
research on microgrids: a review. IEEE Access.
2015;3:890–925.
12 D. KUMAR ET AL.
[4] Luigi F, Nunnari G, Nunnari S. Prediction models for
solar radiation and wind speed time series. In: Fortuna
L, Nunnari G, Nunnari S, editors. Nonlinear modeling
of solar radiation and wind speed time series. Cham:
Springer; 2016. p. 41–46.
[5] Lee DJ, Wang L. Small-signal stability analysis of an
autonomous hybrid renewable energy power genera-
tion/energy storage system part I: time-domain
simulations. IEEE Trans Energy Convers. 2008;23
(1):311–320.
[6] Bansal RC, Bhatti TS, Kothari DP. Some aspects of grid
connected wind electric energy conversion systems.
Interdiscip J Inst Eng. 2001;82(1):25–28.
[7] Adefarati T, Bansal RC. Reliability, economic and
environmental analysis of a microgrid system in the
presence of renewable energy resources. Appl Energy.
2019;236:1089–1114.
[8] Diagne M, David M, Lauret P, et al. Review of solar
irradiance forecasting methods and a proposition for
small-scale insular grids. Renew Sust Energ Rev.
2013;27:65–76.
[9] Kostylev V, Pavlovski A, “Solar power forecasting per-
formance–towards industry standards,” 1st International
workshop on the integration of solar power into power
systems, Aarhus, Denmark, 14 Oct. 2011.
[10] China Energy Storage Network News Center. [cited
2018 Oct 9]. Available from: www.escn.com.cn. (In
Chinese).
[11] De Queiroz AR. Stochastic hydro-thermal scheduling
optimization: an overview. Renew Sust Energ Rev.
2016;62:382–395.
[12] Central Electricity Regulatory Commission (CERC).
2012. New Delhi, Statement of objects and reasons
(terms and conditions for tariff determination from
renewable energy sources). 2019 Jul 25. Available
from: http://www.cercind.gov.in/2012/regulation/RE-
Tariff-Regulations-2012-SOR%206-2-2012.pdf
[13] Achouri F, Mendil B. Wind speed forecasting techni-
ques for maximum power point tracking control in
variable speed wind turbine generator. Int J Model
Simulat. 2019;39(4):246–255.
[14] Zhao X, Liu J, Yu D, et al. One-day-ahead probabilistic
wind speed forecast based on optimized numerical
weather prediction data. Energy Convers Manag.
2018;164:560–569.
[15] Wang J, Song Y, Liu F, et al. Analysis and application of
forecasting models in wind power integration: A review
of multi-step-ahead wind speed forecasting models.
Renew Sust Energ Rev. 2016;60:960–981.
[16] Tascikaraoglu A, Uzunoglu M. A review of combined
approaches for prediction of short-term wind speed
and power. Renew Sust Energ Rev. 2014;34:243–254.
[17] Ahmed A, Khalid M. Multi-step ahead wind forecasting
using nonlinear autoregressive neural networks. Energy
Procedia. 2017;134:192–204.
[18] Lorberfeld A. Machine learning algorithms in layman’s
terms, Part-1. 2019 Mar 2. 2019 Aug 03. Available from:
https://towardsdatascience.com/machine-learning-
algorithms-in-laymans-terms-part-1-d0368d769a7b
[19] Niu T, Wang J, Zhang K, et al. Multi-step-ahead wind
speed forecasting based on optimal feature selection
and a modified bat algorithm with the cognition
strategy. Renewable Energy. 2018;118:213–229.
[20] Yuan X, Tan Q, Lei X, et al. Wind power prediction
using hybrid autoregressive fractionally integrated
moving average and least square support vector
machine. Energy. 2017;129:122–137.
[21] Aydin O, Guldamlasioglu S. Using LSTM networks to
predict engine condition on large scale data processing
framework. Proceedings of the 4th International
Conference on Electrical and Electronic Engineering
(ICEEE), Ankara, Turkey, 2017 Apr 8–10. pp. 281–285.
[22] Srivastava S, Lessmann S. A comparative study of
LSTM neural networks in forecasting day-ahead global
horizontal irradiance with satellite data. Solar Energy.
2018;162:232–247.
[23] Liu H, Mi X, Li Y. Wind speed forecasting method
based on deep learning strategy using empirical wavelet
transform, long short term memory neural network
and Elman neural network. Energy Convers Manage.
2018;156:498–514.
[24] Qu X, Kang X, Zhang C, et al. Short-term prediction of
wind power based on deep long short-term memory.
Proceedings of the IEEE PES Asia-Pacific Power and
Energy Engineering Conference (APPEEC), Xi’an,
China, 2016 Oct 25–28.
[25] Liu H, Mi X, Li Y. Smart multi-step deep learning model
for wind speed forecasting based on variational mode
decomposition, singular spectrum analysis, LSTM network
and ELM. Energy Convers Manage. 2018;159:54–64.
[26] Cao L. Support vector machines experts for time series
forecasting. Neuro comput. 2003;51:321–339.
[27] Kumar J, Goomer R, Singh AK. Long short term mem-
ory recurrent neural network (LSTM-RNN) based
workload forecasting model for cloud datacenters.
Procedia Comput Sci. 2018;125:676–682.
[28] Ban T, Zhang R, Pang S, et al. Referential kNN regres-
sion for financial time series forecasting. International
Conference on Neural Information Processing, Daegu,
Korea: Springer Berlin Heidelberg; 2013 Nov 3. pp.
601–608.
[29] Steftos A. A novel approach for the forecasting of mean
hourly wind speed time series. J Renewable Energy.
2002;27:163–174.
[30] Barbounis TG, Theocharis JB. Locally recurrent neural
networks for wind speed prediction using spatial
correlation. Inf Sci. 2007 Dec 15;177:5775–5797. .
[31] da Cunha Nunes RM. Big data techniques for solar
power forecasting [Dissertation of Master in Data
Analytics and Decision Support Systems]. University
of Porto; 2017.
[32] Bao A, Fei S, Zhong M. Short-term solar irradiance
forecasting using neural network and genetic
algorithm. Chinese Intelligent Systems Conference.
Singapore: Springer, 2016 Oct 22. pp. 619–627.
[33] Kumar R, Aggarwal RK, Sharma JD. Energy analysis of
a building using artificial neural network: A review.
Energy Build. 2013;65:352–358.

[34] Mellit A, Pavan AM. A 24-h forecast of solar irradiance
using artificial neural network: application for perfor-
mance prediction of a grid-connected PV plant at
Trieste, Italy. Solar Energy. 2010;84(5):807–821.
[35] Amjady N, Keynia F, Zareipour H. A new hybrid itera-
tive method for short-term wind speed forecasting. Eur
Trans Electr Power. 2011;21(1):581–595.
[36] Ali M, Khan A, Rehman NU. “Hybrid multiscale wind
speed forecasting based on variational mode decompo-
sition. Int Trans Electr Energy Syst. 2018;28(1):e2466.
[37] Chang GW, Lu HJ, Wang PK, et al. Gaussian mixture
model-based neural network for short-term wind
power forecast. Int Trans Electr Energy Syst. 2017;27
(6):e2320.
[38] Das S, Akella AK. A fuzzy logic-based frequency con-
trol scheme for an isolated AC coupled PV-wind-
battery hybrid system. Int J Model Simulat. 2019;1–13.
DOI:10.1080/02286203.2019.1613063
[39] Madiba T, Bansal RC, Justo J Optimal control system of
under frequency load shedding in microgrid system
with renewable energy resources. In: Smart energy
grid design for Island countries-challenges and oppor-
tunities. Cham, Switzerland: Springer; 2017. p. 71–96.
[40] Guha D, Roy PK, Banerjee S. Grasshopper optimization
algorithm-scaled fractional-order PI-D controller
applied to reduced-order model of load frequency con-
trol system. Int J Model Simulat. 2019;1–26.
DOI:10.1080/02286203.2019.1596727
[41] Mahmoud MS, Abouheaf M, Sharaf A. Reinforcement
learning control approach for autonomous microgrids.
Int J Model Simulat. 2019;1–10. DOI:10.1080/
02286203.2019.1655701
[42] Mishra S, Barisal AK, Babu BC. Invasive weed
optimization-based automatic generation control for
multi-area power systems. Int J Model Simulat.
2019;39(3):190–202.
[43] Li X, Song YJ, Han SB. Frequency control in micro-grid
power system combined with electrolyzer system and
fuzzy PI controller. J Power Sources. 2008;180
(1):468–475.
[44] Bevrani H, Habibi F, Babahajyani P, et al. Intelligent
frequency control in an ac microgrid: online PSO-based
fuzzy tuning approach. IEEE Trans Smart Grid. 2012;3
(4):1935–1944.
[45] Sujil A, Kumar R, Bansal RC, “FCM clustering based
adaptive neuro-fuzzy inference system based PV and
wind generation forecasting agent for energy manage-
ment in a smart microgrid”, 7th IET-RPG conf.,
Copenhagen, Denmark, Sep. 26-27, 2018.
[46] Bansal RC, Pandey JC. Load forecasting using artificial
intelligence techniques: a literature survey.
Int J Comput Appl Technol. 2005;22(2/3):109–119.
[47] Sonker B, Kumar D, Samuel P. Design of two degree of
freedom-internal model control configuration for load
INTERNATIONAL JOURNAL OF MODELLING AND SIMULATION 13
frequency control using model approximation.
Int J Model Simulat. 2019;39(1):27–37.
[48] Acharyulu BS, Mohanty B, Hota PK. Performance ana-
lysis of a cascade controller for conventional and
deregulated power systems. Int J Model Simulat.
2019;1–19. DOI:10.1080/02286203.2019.1610690
[49] Kumar D, Mathur HD, Bhanot S, et al. Frequency
regulation in islanded microgrid considering stochastic
model of wind and PV. Int Trans Electr Energy Syst.
2019;29(9):e12049.
[50] Long Short-Term Memory (LSTM). 2019 May 30.
https://in.mathworks.com/discovery/lstm.html
[51] Flores JJ, Graff M, Rodriguez H. Evolutive design of
ARMA and ANN models for time series forecasting.
Renewable Energy. 2012;44:225–230.
[52] Laopaiboon T, Ongsakul W, Panyainkaew P, et al.
Hour-ahead solar forecasting program using back pro-
pagation artificial neural network. IEEE International
Conference and Utility Exhibition on Green Energy for
Sustainable Development, 2018 Oct 24; Phaket,
Thailand. pp. 1–7
[53] Syed MS, Suresh CV, Sivanagaraju S. Short term solar
insolation prediction: P-ELM approach. Int J Parallel
Emergent Distrib Syst. 2018;33(6):663–674.
[54] Tena García JL, Cadenas Calderón E, González
Ávalos G, et al. Forecast of daily output energy of
wind turbine using sARIMA and nonlinear autoregres-
sive models. Adv Mech Eng. 2019;11(2):1–15.
[55] Cadenas E, Rivera W, Campos-Amezcua R, et al. Wind
speed prediction using a univariate ARIMA model and
a multivariate NARX model. Energies. 2016;9(2):109.
[56] Cadenas E, Rivera W. Wind speed forecasting in three
different regions of Mexico, using a hybrid ARIMA–
ANN model. Renewable Energy. 2010;35
(12):2732–2740.
[57] Nakagawa S, Tokoro T, Nakano T, et al. An effect of
snow for electric energy generation by 40 kW PV
system. IEEE Proc. 3rd World Conference on
Photovoltaic Energy Conversion, 2003 May; Osaka,
Japan. vol. 3, pp. 2447–2450.
[58] Matsuo H, Kobayashi K, Sekine Y, et al. Novel solar cell
power supply system using the multiple-input DC-DC
converter. INTELEC - Twentieth International
Telecommunications Energy Conference, IEEE San
Francisco, CA, USA, 1998 Oct 4-8. pp. 797–8022.
[59] Shimizu T, Hirakata M, Kamezawa T, et al. Generation
control circuit for photovoltaic modules. IEEE Trans
Power Electron. 2001;16(3):293–300.
[60] Senjyu T, Nakaji T, Uezato K, et al. A hybrid power system
using alternative energy facilities in isolated island. IEEE
Trans Energy Convers. 2005;20(2):406–414.
[61] Bevrani H, Feizi MR, Ataee S. Robust Frequency Control
in an Islanded Micro-grid: and μ -Synthesis Approaches.
IEEE Trans Smart Grid. 2016;7(2):706–717.