Journal of Reliable Intelligent Environments (2023) 9:183–200

https://doi.org/10.1007/s40860-021-00166-x

ORIGINAL ARTICLE

A comprehensive study and performance analysis of deep neural

network-based approaches in wind time-series forecasting
Md. Mijanur Rahman1 · Mohammad Shakeri3 · Fatema Khatun2 · Sieh Kiong Tiong3 ·
Ammar Ahmed Alkahtani3 · Nurul Asma Samsudin3 · Nowshad Amin3 · Jagadeesh Pasupuleti3 ·
Mohammad Kamrul Hasan4

Received: 11 May 2021 / Accepted: 15 December 2021 / Published online: 11 January 2022
© The Author(s), under exclusive licence to Springer Nature Switzerland AG 2022

Abstract
The increasing energy demand and expansion of power plants are provoking the effects of greenhouse gas emissions and
global warming. To mitigate these issues, renewable energies (like solar, wind, and hydropower) are blessings for modern
energy sectors. The study focuses on wind-speed prediction in energy forecasting applications. This paper is a comprehensive
review of deep neural network based approaches, like the “nonlinear autoregressive exogenous inputs (NARX)”, “nonlinear
input-output (NIO)” and “nonlinear autoregressive (NAR)” neural network models, in time-series forecasting applications.
This study proposed NARX based prediction models in wind-speed forecasting for short-term scheme. The meteorological
parameters related to wind time-series have been analyzed, and used for evaluating the performance of the proposed models.
The experiments revealed the best performance of the prediction models in terms of “mean square error (MSE)”, “correlation-
coefficient (R 2 )”, “auto-correlation”, “error-histogram”, and “input-error cross-correlation”. Comparing with the other neural
network models, like “recurrent neural network (RNN)” and “curve fitting neural network (CFNN)” models, the NARX-
based prediction model achieved better performance in regard to “auto-correlation”, “error-histogram”, “input-error cross-
correlation”, and training time. The results also showed that the RNN and CFNN models performed better prediction accuracy
with R 2 and MSE values. While this performance index is slightly higher, it is negligible in forecasting applications and
concluded that the proposed NARX-based model achieved the better prediction accuracy in terms of other performance
evaluation measures.

Keywords NARX neural network · Recurrent neural network · Renewable energy · Time-series forecasting · Wind-speed
prediction

B Md. Mijanur Rahman

[email protected]
Mohammad Shakeri Mohammad Kamrul Hasan
[email protected] [email protected]
Fatema Khatun 1 Department of Computer Science and Engineering, Jatiya
[email protected] Kabi Kazi Nazrul Islam University, Trishal, Mymensingh
Sieh Kiong Tiong 2224, Bangladesh
[email protected] 2 Department of Electrical and Electronic Engineering,
Ammar Ahmed Alkahtani Bangabandhu Sheikh Mujibur Rahman Science &
[email protected] Technology University, Gopalganj, Dhaka 8100, Bangladesh
3 Institute of Sustainable Energy, Universiti Tenaga Nasional
Nurul Asma Samsudin
[email protected] (The National Energy University), Jalan Ikram-Uniten, 43000
Kajang, Selangor, Malaysia
Nowshad Amin
4 Center for Cyber Security, School of Information Science and
[email protected]
Technology, Universiti Kebangsaan Malaysia UKM, 43600
Jagadeesh Pasupuleti Bangi, Selangor, Malaysia
[email protected]

123
184 Journal of Reliable Intelligent Environments (2023) 9:183–200

1 Introduction crucial role in the wind-energy production (WEP), and opti-

mum planning of WT installation. The short-term prediction
The global electricity or energy demand is increasing day scheme helps to control the power system [9] and ensure
by day. This leads to rising greenhouse gas emissions due to the WT operation and ranged between a day ahead till 30
the high consumption of energy produced by fuel generators. min, respectively [10]. But, the prediction is a very challeng-
Fossil-based sources can be reduced by utilizing renewable ing task as the wind time-series related to many factors, like
energies, such as wind-power, solar-energy, biomass, etc. wind-speed and direction, solar-irradiance, weather temper-
Among them, the abundant and pollution-free renewable ature and pressure, rotation of the earth, etc. [11,12].
nature of wind is a rising energy resource in global energy Recently, many researchers are conducting their works in
consumption [1]. According to the “EU Renewable Energy designing forecasting models for renewable energies. The
Directive”, the renewable energy share in global energy con- forecasting approaches are commonly divided into three cat-
sumption has increased from 7.4 percent in 2018 to 8.7 egories; physical, statistical and hybrid models [13,14]. The
percent in 2019 [2], as shown in Fig. 1, and the worldwide physical approaches depend on “numerical weather predic-
cumulative volume of installed wind power reached about tion (NWP)” that use weather forecast data (like wind-speed
650.8 GW in 2019 [3], in which China has the major contri- and direction, temperature, pressure, etc.) to estimate the
bution with a capacity of over 237 GW [4], as illustrated in output value. Additionally, it includes information about
Fig. 2. The survey report of the “World Wind Energy Asso- the physical characteristics of the wind-turbine and its sur-
ciation (WWEA)” stated that about 12% of the worldwide rounding. Thus, these models gather weather-related and
electric energy requirements is expected through wind power
by the end of the year 2020 [5,6]. But the wind turbines
continuously need sufficient wind for the power generation
which mainly depends on meteorological conditions. Thus,
energy generation from wind plants faces difficulties due to
the uncertainties of data origin. Hence, research is being car-
ried out to address this unpredictability and develop accurate
wind-speed forecasting models [7] in the future.
The most considering thing about the wind-energy gen-
eration is to choose the stationary location for installing
wind-turbine (WT) that produces maximum energy produc-
tion. Since the wind-energy is proportional to the cubic
relation of wind-speed, any minor variation in wind-speed
can cause a huge difference in energy production. Hence,
the estimation of wind-speed can assist to reduce the risks
of investment in installing wind-turbines in different places
[8]. In other words, wind-speed prediction (WSP) has a

Fig. 2 The volume of cumulative installed wind-power capacity in the

Fig. 1 The global share of renewable energy in total consumption [2] globe

123
Journal of Reliable Intelligent Environments (2023) 9:183–200 185

forecast wind-power from the available data at a typical wind-

farm.
One of the most promising AI models is the “artifi-
cial neural network (ANN)” that has been effectively used
in forecasting, pattern recognition, and energy assessment
applications [31]. It is capable of self-learning and self-
adaption for training of past observations in time-series
prediction problems [32–34]. Most of the ANN-based pre-
diction models include “back-propagation neural network
(BPNN)”, “recurrent neural network (RNN)”, and “radial
basis function neural network (RBFNN)”. The BPNN [35]
simplifies the multilayer-feedforward network to predict the
NWP, and the RNN has a potential to achieve better fore-
Fig. 3 Basic modules of the statistical approach in time-series forecast- cast outcomes compared to other approaches [36–38]. These
ing
statistical approaches have been effectively utilized in both
wind-speed estimation [39] and wind-power prediction [40].
The curve-fitting is a method of regression analysis that
environmental or physical data as inputs and estimate wind- presents the best fit to the particular curves in the time-series
speed accordingly for the wind-energy production. Another dataset. The “curve-fitting neural network (CFNN)” was used
useful information is the theoretical wind speed-to-power to estimate wind-speed as well as wind-power previously in
curve (see Sect. 2.1) that indicates expected wind-power to [41].
be produced for wind-speed at a typical wind-farm [15–17]. Fadare [42] proposed a model to predict wind-speed
Prediktor [18] is a physical model that used NWP in order to using ANN method that utilized the gradient descent back-
convert wind data through a power curve to yield the forecast propagation algorithm [43] for training the model. The model
value. received the meteorological parameters as inputs and pro-
The statistical methods [19] mostly rely on a huge past data duced the estimated wind-speed as outcomes. Filik et al.
and their statistical relations without considerable physical [44] proposed another ANN-based model that used several
phenomena, and use “supervisory control and data acqui- climatological quantities and wind-speed, pressure and tem-
sition (SCADA)” [20] and “numerical weather prediction perature data for wind-speed predictions in various cases. A
(NWP)” data to forecast the final outcomes, as shown in combined method [45] of “artificial neural network (ANN)”
Fig. 3 [21]. NWP data contains the values of wind-speed and “support vector machine (SVM)” was presented in
and direction, weather temperature and humidity, etc., at the a short-term scheme with different time-steps that used
region where the WTs are installed. These models can fore- SCADA system with weather and physical data in wind-
cast different time horizons’ values from seconds to hours and power forecasting. The authors [46] proposed the same type
originate functional dependencies directly from past obser- of prediction method based on ANN that used historical data
vations. Hence, they are also known as data-driven models. provided by the SCADA system without exogenous input.
The statistical methods, including artificial intelligence or Another short-term prediction model was proposed in [30]
machine learning and time-series analysis approaches, have that utilized the machine-learning method for feature selec-
been recently implemented in forecasting applications [22– tion of wind data and regression process in power forecasting.
24]. Previento [25] is a similar model to Prediktor [18], but Jursa et al. [47] introduced a prediction model that combined
with local prediction and uncertainty measurement. This sta- the nearest neighbor search with the neural network in power
tistical model used a wind-power prediction tool [26] that prediction for a short-term horizon.
considers both the dynamic behavior of wind-speed and As the wind-speed has an instability nature over time,
direction in the power generation. it is hard to predict accurate energy in the short-term hori-
Generally, hybrid models combine the physical and statis- zon. Thus, several advanced approaches have been combined
tical methods [27,28], and mostly used in weather forecasting linear and nonlinear methods to govern external inputs in
and time-series analysis. The advanced prediction mod- improving short-term predictions based on multistep RNN
els can be implemented using “artificial intelligence (AI)” [48,49], “multivariable long short-term memory (LSTM)”
including machine learning, neural networks, neuro-fuzzy, [50,51], “autoregressive moving average (ARMA)”, and
and some other hybrid models [29,30]. These advanced “nonlinear autoregressive exogenous input (NARX)” models
approaches can show significant accuracy in wind-power [52,53]. Very few works have been done in time-series fore-
prediction for a short-term scheme. Zephyr [27] is a hybrid casting using these hybrid models. The study [54] presented
model that combined both the statistical and NWP tools to a single-step and a multistep RNN model that predicted daily

123
186 Journal of Reliable Intelligent Environments (2023) 9:183–200

meteorological data. It is expressed as follows:

 
X i = Ospeed ; Otemp ; Odirection ; Opressure ; . . . (1)

This research aims to estimate wind-power, but the wind-

power is produced from the kinetic energy of air in the wind-
turbines, and this kinetic energy is transformed to power by
using the following equations [64,65]:

1
Pavail = ρ Av3 CP (2)
2

where v = air speed (m/s), ρ = air density (kg/m3 ), A =

Fig. 4 Illustration of wind speed-to-power curve in terms of theoretical turbine-blade area (m2 ), Cp = power coefficient.
and practical values [68]
From the power equation, it was proven that 20 percent
increasing in wind-speed will rise 73 percent of wind-power
production [66]. Another better option is to acquire wind-
to monthly wind-speed estimated values. A multistep ahead power from the speed-power plot [67] by using estimated
approach [55] was applied in short-term forecasting of solar- wind-speed data, and subsequently, this paper highlights the
irradiation, where the temperature was chosen as an external wind-speed prediction. A typical wind speed-to-power plot
input in the NARX neural network model. with theoretical and practical values is illustrated in Fig. 4.
Considering wind data as a time-series and improve- The first region is referred to as the cut-in value, where the
ment of dynamic neural networks in time-series forecasting speed value does not satisfy the thresholding and provides the
[56,57], the NARX neural network is presented in this paper power output “zero”. The second region up to the rated speed
for predicting wind time-series with a short-term basis. It is provides the growing of generated power. Lastly, the third
a regression of nonlinear external inputs that is determined area until the cut-off speed reaches the rated power-output.
by the nonlinear dynamic model [58]. The NARX models After reaching the cut-off value, the WT will be off to keep
can be applied in many problem domains, such as future its parts safe from high breezes and then the power-output
value prediction [59], time-series modeling [60] classifica- will be “zero”.
tion problems [61], pattern recognition [62], and intelligent Naturally, winds hold the features of non-linearity and
system design [63]. This study highlights the nonlinear nature non-stationarity [69]. Thus, the wind speed-to-power trans-
of wind time-series and investigates the ability of NARX formation is also non-linear that yields uncertainty in wind-
model in forecasting problems. The proposed NARX based power generation [68,70], and practically, it depends on many
approach utilized supplementary exogenous variables that is factors. While a linear model is applied in a non-linear prob-
carried out to estimate wind-speed for short-term scheme. lem, it affects the prediction output, and hence, this study
This study also focuses the performance analysis of the highlights the non-linear modeling of prediction approach
NARX model in wind-speed prediction, and investigates the with wind time-series. In a prediction problem (see Fig. 5),
performance measures of the NARX model with other ANN- a predicted value Ŷ is a generated time-series that can be
based models. obtained from the past observations of the wind time-series
and meteorological data.
2 Wind time-series and NARX model It has two dimensions, such as a time-horizon H (the time-
steps ahead in the future) and a number of predictions. Thus,
2.1 Time-series forecasting it could be a single-step or multi-step prediction, and repre-
sented as follows:
Generally, a time-series input X is represented by a col-  
lection of observations with a particular time-stamp t, i.e., Ŷ = ŷn+1 ; ŷn+2 ; . . . ; ŷn+H (3)
X = [x1 , x2 , . . . , xt ] of a collection of time-dependent
variables. The wind-turbines consist of several sensors that In wind prediction problems, the time-horizon is divided into
offer real-time climatological data about the environment three main categories [71,72]; these are (1) very short-term
(like wind-speed/direction, pressure, humidity, temperature, (about few hours), (2) short-term (day long), and (3) long-
etc.) to be analyzed to produce wind-energy. Thus, a wind term (few days or week) forecasts, with time-step of seconds
time-series includes observations over several years that con- to days, weeks or months. For short-term scheme, machine
taining several variables related to wind, wind-energy, and learning models are applied in wind-speed and wind-power

123
Journal of Reliable Intelligent Environments (2023) 9:183–200 187

NN model [81]. This network utilizes its memory features to

provide full performances for nonlinear time-series predic-
tion using the past true time-series or past predicted values.
The typical structure of NARX-NN is given in Fig. 6c, which
involves external inputs and true-outputs. The predicted out-
put (at p time-ahead) is expressed as follows:

y(τ + p) = F{x(τ ), x(τ − 1), . . . , x(τ − dx ),

(4)
y(τ ), y(τ − 1), . . . , y(τ − d y )}

Where x(τ ) is the input and y(τ ) is the output at time τ , dx

Fig. 5 Illustration of inputs and estimated outputs in wind time-series is the input-delay, and d y is the output-delay. The past values
prediction of the level dx and d y are also known as “tapped delay lines
(TDLs)” [82] which stock the past observations. Thus, the
predicted outputs rely not only on the present and past inde-
prediction, and provide better accuracy in prediction results
pendent (exogenous) inputs, but also on the previous output
[73,74]. “ANN (artificial neural network)”, “ARMA (autore-
of the network. The dynamic and nonlinearity characteris-
gressive moving average)” and “ARX (autoregressive exoge-
tics of this method plus the delayed input-output or feedback
nous)” models are commonly used in the short-term or very
outputs allow the computation as well as determination tasks
short-term scheme. The short-term and medium-term fore-
in time-series prediction problems that are almost impossible
casting schemes usually incorporate all the methods used for
to solve by conventional linear methods. Thus, this advanced
shorter predictions and physical characteristics [21]. This
model perfectly fits in wind time-series prediction with better
study presents the deep neural network based wind-speed
accuracy.
prediction approach that requires a set of observations from
meteorological dataset. Wind-speed is the true-output vari-
able and all others available in dataset (like temperature, Learning and training modes in NARX-NN
humidity, density, etc.) are the auxiliary variables used for
training the model. The learning method used in the NARX-NN model is more
efficient than other neural networks in non-linear systems
2.2 NARX neural network models [83]. The model uses faster network connection with good
generalization ability and the improved gradient-descent
The models based on ANN approaches are commonly learning algorithm that makes the network more active. There
divided into two groups, static models and dynamic mod- are two different modes of NARX-NN training, (1) series-
els. The dynamic neural networks are further classified into parallel mode (also refer to open-loop) and (2) parallel mode
feedback and no-feedback types [75]. The dynamic neural (also refer to close-loop), as shown in Fig. 7 [82,84]. In
networks with feedback-loop are commonly applied in time- open-loop mode, output ŷ(τ ) is measured from current and
series forecasting applications, in which the forecast values previous values of external input x(τ ) as well as the true-
depend on past and present inputs as well as the estimated out- output y(τ ) (also known as the desired output). Whereas, in
put [76]. Moreover, the time-series predictions using ANN the close-loop, the prediction value ŷ(τ ) is estimated from
approaches can be categorized into three different nonlin- the previous output and the external input that are the same
ear problems [77]. These are the “nonlinear input-output as in open-loop mode. Thus, the time-series prediction using
neural network (NIO-NN)”, the “nonlinear autoregressive NARX-NN model is performed in the following modes:
neural network (NAR-NN)”, and the “nonlinear autoregres-
sive exogenous neural network (NARX-NN)”, as illustrated 1. Training mode: As the true outputs are available during
in Fig. 6 [78]. training phase, the series-parallel or open-loop configura-
The NARX-NN model is a feedback dynamic neural net- tion is used. The true-output y(t) is fed to the model with
work that is used for prediction purposes in this study. This the external input x(t). At the time, the output feedback
advanced model is distinguished from the BP network by is detached; so, the resulting network is performed as an
considering the previous values of external variables, and the elementary multi-layer feedforward fashion. The well-
feedback property of the NARX-NN makes it a special sort of established standard back-propagation algorithm can be
RNN [79] that has a potential in time-series forecasting [80]. used for training multi-layer perceptron (MLP) networks
Unlike a conventional RNN, the output is fed back from the and this configuration permits to train any sort of contin-
output-neurons, not from the hidden-neurons in the NARX- uous nonlinear-mapping datasets [57].

123
188 Journal of Reliable Intelligent Environments (2023) 9:183–200

Fig. 6 The typical structures of different neural networks [78]

Fig. 7 Two configurations of the NARX-NN model in time-series forecasting

123
Journal of Reliable Intelligent Environments (2023) 9:183–200 189

(a)

(b)

(c)

Fig. 9 Patterns of wind-speed and other meteorological data

Fig. 8 NARX-ANN-based methodological steps of the wind time- tological data constraints. The preprocessed data cannot
series forecasting be directly fed to the model, the normalization process
is required to transform the data into a form suitable to
feed the neural network.
2. Operational mode: After training, the true-output line is
3. Network design and selection: Different advanced neural
detached and replaced with the feedback output ŷ(τ ).
networks are created to investigate multiple meteorolog-
Hence, the NARX structure is converted to the close-
ical data effect. The suitable structure of the prediction
loop configuration that is beneficial for estimating future
model is trained with available data for forecasting.
values [85]. In this research, the NARX-NN structures
4. Training, validation and testing: The training of the
were created for performing the wind-speed prediction.
selected model involves the modification of weights in
a planned and ordered way using an appropriate learn-
ing algorithm to adopt the network parameters. After
training, validation is required to evaluate the model per-
3 Materials and methodology formance while tuning the model’s hyper-parameters and
lastly, testing is accomplished on unknown dataset to esti-
The NARX-NN based prediction model is presented in this
mate future values.
study, and applied successfully in the wind time-series fore-
casting. The methodology for wind data prediction is outlined
in Fig. 8 that comprises the following steps:
3.1 Data description
1. Data collection and preprocessing: This step involves col-
lecting wind-related data from more extensive databases The energy generation depends on various weather-related
and different data sources. The preprocessing including data constraints, including wind-speed and direction, pres-
data enrichment as well as cleaning that can deal with sure, temperature, humidity, etc. [86]. In a HRES (“hybrid
noisy and missing information. renewable energy system”), the records of meteorological
2. Data filtering and transformation: The irrelevant informa- values are received from a weather station and/or a wind-
tion makes the network training more difficult. Filtering farm. In this research, the wind dataset with meteorological
is applied to select useful features within several clima- values have been obtained from the “National Renewable

123
190 Journal of Reliable Intelligent Environments (2023) 9:183–200

Table 1 The distributed wind resource with some descriptive statistic

Distributed wind resource (2007–2013)

Pressure Temperature Wind speed Wind direction

Average annual pressure @ Average annual temperature Average annual wind speed Average annual wind direc-
0-Meter Hub height @ 40-Meter Hub height @ 40-Meter Hub height tion @ 40-Meter Hub height

Year Samples Year Samples Year Samples Year Samples

2007 10,00,000 2007 10,00,000 2007 10,00,000 2007 10,00,000

2008 10,00,000 2008 10,00,000 2008 10,00,000 2008 10,00,000
2009 10,00,000 2009 10,00,000 2009 10,00,000 2009 10,00,000
2010 10,00,000 2010 10,00,000 2010 10,00,000 2010 10,00,000
2011 10,00,000 2011 10,00,000 2011 10,00,000 2011 10,00,000
2012 10,00,000 2012 10,00,000 2012 10,00,000 2012 10,00,000
2013 10.00.000 2013 10.00.000 2013 10.00.000 2013 10.00.000

Energy Laboratory (NREL)” of the “U.S. Department of and degrades the forecasting performance. The filtering is
Energy” [87]. The distributed wind resource consists of 7 required to choose only useful features from meteorologi-
years (2007-2013) data set of pressure, temperature, wind- cal inputs that improve prediction performance and reduces
speed, and wind direction. The average annual meterologcal storage space. Therefore, two filters, irrelevancy filter and
data in the resource were acquired at different surface levels, redundancy filter are applied on the input vectors for extract-
such as pressure at 0-Meter, the temperature at 40-Meter, ing useful features [30,88].
wind-speed at 40-Meter, and wind direction at 40-Meter The irrelevancy filter computes “mutual information
above surface level. Each data set contains 1 (one) million (MI)” values through “binomial distribution” among the tar-
data samples. The outlines of the data description with some get outputs and the external inputs [89]. Then the values of the
descriptive statistics are given in Table 1 The pattern of the variables are selected based on the MI values and a specific
wind time-series (speed, pressure, and temperature) is illus- threshold (T1 ). The variables owning MI values larger than
trated with graphs in Fig. 9. The time-interval τ is chosen as T1 are received, and other values are worthless. The thresh-
the number of days, hours, or minutes, depends on the avail- old value is dynamically adjusted based on input variables,
ability of datasets; wind data are available in this dataset which is different for different meteorological parameters.
with minutes-interval. The output is determined as a nonlin- The filter method uses the proper threshold value (trial and
ear function of the time-series and their past values. Suppose error basis) to select the most relevant features. This filter
X S (τ ), X T (τ ), and X P (τ ) denote wind-speed, temperature, delivers a feature set, say F1 (t), which contains relevant fea-
and pressure at time τ , respectively. Thus, the estimated out- tures, then passes to the redundant filter. But this feature set
put Y (τ ) is expressed as follows: may carry redundant data that can degrade the forecasting
performance. Therefore, the redundancy filter is required to
Y (τ ) = [X S (τ ), X T (τ ), X P (τ ), Y (τ − 1)] (5) discard the redundant candidates from F1 (t).
The redundant candidates can be defined on the basis
The external input variables contain more values so that no of the enormous MI value between two picked candidates.
useful information in the dataset is missed. The data cannot Thus, more mutual information indicates a higher redun-
be directly fed to the network model, it is required to be dancy score. The “redundancy criterion (RC)” equation is
normalized and reduced to useful features to achieve accurate used to quantify the redundancy score of each chosen candi-
output. date with others, given as-
 
Data filtering and transformation RC[xi (t)] = max x j (t)∈F1 (t)−xi (t) MI[xi (t), x j (t)] (6)

The prediction performance is dependent on multiple mete- The candidate features of F1 (t) is then ranked according
orological variables (such as wind-speed, direction, temper- to the above redundancy measure. A large RC[xi (t)] value
ature, and pressure). There is a nonlinear mapping relation implies xi (t) is a higher redundant piece or, equivalently, a
among these variables and their past values. Thus, a vast input less informative candidate. If RC[xi (t)] > T2 (redundancy
dataset includes ineffective features (such as irrelevant and threshold), xi (t) is a redundant candidate. So, among the can-
redundant candidates) that make the training more difficult didate xi (t) and its rival, say xr (t), one will be eliminated. In

123
Journal of Reliable Intelligent Environments (2023) 9:183–200 191

Fig. 10 Two different prediction model’s architectures based on external inputs

this filtering approach, the elimination is done based MI val- 3.2 Prediction models
ues between them (i.e., MI[xi (t), X (t)] and MI[xr (t), X (t)]
[37]. The candidate feature owning a smaller MI value is The proposed prediction approach for wind-speed fore-
removed from the relevant set, F1 (t). Finally, this redundant casting is employed using different NARX neural network
filter delivers a feature set, say F2 (t), including relevant and structures that investigate several climatological data effects.
non-redundant candidate inputs. Although there are multiple The network architectures with two different meteorological
features associated with input variables, only a few useful parameters, presented in Fig. 10, are examined in this study.
features are left in F2 (t) and they are used in time-series The first model involves only one input (i.e., wind-speed),
forecasting. and the second model uses other meteorological parameters
In addition, several studies proposed correlation analysis (temperature and pressure values) along with wind-speed.
in forecasting problems [86,90]. However, wind time-series In this study, the prediction model based on NARX neu-
hold a nonlinear mapping association among many input ral network has been applied in the short-term scheme (i.e.,
variables, and correlation analysis is a linear feature deter- hourly-basis speed prediction with a certain time-ahead). The
mination procedure. Thus, it may not precisely determine designated model requires two types of entries, such as the
the features of the input variables in time-series predic- actual past values (the desired output) and several external
tion. Besides, correlation analysis simply views the relevancy (exogenous) inputs. As the target value is wind-speed, the
between the desired outputs and external inputs. endogenous input is wind-speed as well as temperature and
The data have been normalized first, before feeding to the pressure are being considered as the exogenous variable.
network model. Normalization transforms the data values of When the data from the external inputs x(τ ) is involved
different ranges into identical forms. The normalized data for measuring the output y(τ ), the NARX model attains
meets the requirements of the adaption of the input neurons improvement in prediction applications and affords perfect
that makes the training process faster. In the final phase, the output compared to other ANN based approaches.
output is denormalized to obtain the actual estimated values. While the proposed model is applied in the short-term pre-
The process of normalization is done by mapping prepos- diction scheme and the information of upcoming exogenous
sessed data into a matrix-row with the values of [− 1, + 1], inputs is available, multi-step ahead forecasts are possible to
using the following formula: achieve. It is done through an iterative process in which the
output of one-step forecasting is consumed as the input in
x − xmin the succeeding stage, as illustrated in Fig. 11.
xnorm = 2.0 ∗ − 1.0 (7)
xmax − xmin

123
192 Journal of Reliable Intelligent Environments (2023) 9:183–200

Fig. 11 Process of multi-step ahead time-series predictions

This lets the input time-series remain unchanged and eas-

ily adapt it to the network model with several time-delays.
The suggested model using NARX neural network approach
includes three neurons (two for external inputs and one for
targets) in the input-layer, one hidden-layer, one neuron (for
forecast value) in the output-layer, and two tapped-delay-
lines (TDLs). TDLs store the past values of input x(τ ) and
output y(τ ). The output layer has only one neuron that pro-
vides one predicted value (wind-speed) at a time, and this
Fig. 12 The flowchart of the proposed prediction model design and
prediction output is fed back to the input. Generally, there is
training process
no standard rule to fix the number of neurons in the hidden-
layer. This parameter is convinced by trial and error. More The error is measured as follows:
neurons leading to raise the power of the network model,
1 
N
but it involves additional computation, and perhaps causes 2
over-fitting. The input and output neurons require a lin- E(τ ) = yk (τ ) − ŷk (τ ) (8)
2
ear activation function, whereas the hidden neurons use the k=1

sigmoidal activation function in input-to-output data progres-

sion. where yk (τ ) and ŷk (τ ) are the true and measured values at
time τ , respectively. The tanh transfer function is applied
in hidden-layer to activate the neurons in the network, as
follows:
Training process
a 
i
y = f (ai ) = tanh (9)
The proposed network model is primarily designed as multi- 2
layer feedforward fashion. The network’s learning and train-
ing process involves the modifications or adjustments of the Where ai is total inputs of the i neuron. The neuron’s weight-
weights (i.e., network parameters) using a suitable supervised updating between the layer i and j, is performed as follows:
learning algorithm. Hence, an optimization technique with
gradient-descent is employed to reduce the errors between wi j (n + 1) = wi j (n) + wi j (n) (10)
the true-outputs and the estimated-outputs [91,92]. It helps
to yield the forecast-value closer to the target-output as well Where wi j (n) is the required weight-variation in nth rep-
as to avoid over-training. In the training of the model with etitions, calculated from the negative error-gradient with the
the gradient-descent technique, the weight and bias values network learning rate η, as follows:
are successively updated by error back-propagation (EBP)  
algorithm that minimizes the combination of squared errors. wi j (n, t) = −η × (∂ E(n, t))/ ∂wi j (n) (11)

123
Journal of Reliable Intelligent Environments (2023) 9:183–200 193

The flowchart of the overall prediction model design and Table 2 Prediction performances for considering only one input and
training process using NARX neural network paradigm is all available external inputs
presented in Fig. 12. During training, the external inputs Input H&D MSE R2 Training time
and the true outputs are declared to the network model.
Temperature 20H and 1D 4.57256 e−4 9.2911 e−1 10 s
The network models are created as autoregressive, in which
only inputs are lagged desired values and lagged exogeneous Air-pressure 20H and 1D 4.79607 e−4 9.2408 e−1 06 s
input values, and hence, the output-feedback is disconnected Wind speed 20H and 1D 4.50001e−5 9.97395 e−1 03 s
(i.e., open-loop mode). Once the network is trained, the All 20H and 1D 1.30786 e−5 9.99860 e−1 20 s
output-feedback connection is established, and thus, the
trained network is transformed from series-parallel to par- Table 3 Prediction performance of the proposed NARX-NN model
allel configuration (i.e., close-loop mode). In this stage, the with different hidden-neurons (H) and delays (D)
measured output ŷ(t) is fed back as input with other external
H&D MSE R2 Training time
inputs to estimate the final forecast values. The advanced
“Levenberg-Marquardt back-propagation algorithm” [93] 22H and 2D 1.97256 e−4 9.97911 e−1 14 s
has been applied in training the network because it is by far 20H and 2D 1.49607 e−4 9.98408 e−1 17 s
the most popular advanced supervised learning algorithm. 19H and 1D 1.55916 e−4 9.98345 e−1 10 s
The Bayesian Regularization is another improved back- 20H and 1D 1.30786 e−5 9.99860 e−1 20 s
propagation algorithm that can be used also in non-linear 10H and 2D 1.51361 e−4 9.98387 e−1 11 s
problems. The model’s prediction performance is evaluated 10H and 1D 1.61457 e−4 9.98285 e−1 12 s
using the metrics, MSE (“mean square error”) and R 2 (“cor-
relation coefficient”) formulas.
in this research work. The “number of hidden layers” and the
“number of neurons in the hidden layers” are two essential
4 Results and discussion parameters in designing deep neural network models. In this
study, the ANN-based prediction models have been devel-
All the required WSP data were taken from the US “National- oped with different hidden layers (for example, the NARX
Renewable-Energy-Laboratory (NREL)” wind dataset for network model in Fig. 13 has 20 hidden layers). The com-
short-term prediction [87]. Ten thousand samples of wind parative performance analysis of these prediction models is
time-series (hour/year) were received for experiments, in presented in Tables 2, 3, 4, and 5, regarding the number of
which 80% training, 10% testing, and 10% validation sam- hidden layers (H) and feedback delays (D).
ples. This study investigated the NARX-NN model’s perfor- This study concerns the better understanding of ANN-
mance compared to NIO-NN, NAR-NN, RNN and CFNN based approaches and learning algorithms to reduce the
models in wind time-series forecasting. Figure 13 shows the computational complexity of the learning process, improve
structures of the different types of network models developed forecasting performance, and reduce the space complexity in

Fig. 13 The developed ANN-based models for wind-speed prediction,

123
194 Journal of Reliable Intelligent Environments (2023) 9:183–200

addressing wind time-series. The dataset consists of selected air-pressure) for better forecasting accuracy in this study.
features that were used as inputs for the proposed models to The prediction performances of the proposed NARX model
predict wind-speed. for considering only one external input (wind-speed or tem-
All the models have been developed in the MATLAB plat- perature or pressure) and all available external inputs are
form with Deep Learning Toolbox and demonstrated their separately presented in Table 2.
results under the same conditions. The scripts were gener- The experiments examined the model’s performance in
ated using the NTSTOOL (neural network time series tool) terms of R 2 and MSE values. The R 2 identifies the rela-
functions. The ’tonndata’ function was used to transform time tionship within measured and predicted values which ranged
series data into standard neural network cell array form. The between 0 (zero) and 1 (one). The R 2 value closer to 1
NARX model was created using the ’narxnet’ function with (one) means a better fit of the network model. On the other
’inputDelays’, feedbackDelays and hiddenLayerSize param- hand, the MSE value shows the distinction between the true-
eters. The NIO model was implemented by a time-delay outputs and the target-outputs on average squared; it’s lower
neural network and created using the function ’timedelaynet’ value means better prediction accuracy in the experiments.
with ’inputDelays’ and ’hiddenLayerSize’ parameters. The Using all external inputs provides better prediction perfor-
NAR model was created using ’narnet’, and the RNN model mance (MSE = 1.30786e−5 and R 2 = 9.99860e−1 ), but it
was created using ’layrecnet’ function with ’feedbackDe- is required more computation as well as more training time (=
lays’ and ’hiddenLayerSize’ parameters. The CFNN model 20 s). Considering all external inputs (temperature, pressure
was created using the function ’fitnet’ with ’hiddenLay- and wind-speed) in the proposed model, Table 3 illustrates
erSize’ parameter. The function ’preparets’ prepared time the WSP performance with several hidden-neurons (H) and
series data for training and simulation in a particular network. applied delays (D).
The function ’dividerand’ randomly assigned target values to According to Table 3, the proposed NARX-based pre-
training, validation, and test sets during training. The ’train’ diction model presented the best performance (R 2 =
function used ’trainlm’ (Levenberg-Marquardt backpropa- 9.99860e−1 and MSE = 1.30786e−5 ) with H = 20 and
gation) algorithm for training the network model. We have D = 1. This best performance was used for comparing the
customized performance parameters at ’net.performParam’ proposed NARX based model with other ANN-based mod-
and plot parameters at ’net.plotParam’. Finally, the network els (like NAR, NIO, RNN, CFNN) in terms of MSE and
models were tested using the functions’ net’ (outputs of the R 2 values, epochs, training time as well as other network
network), ’gsubtract’ (errors between outputs and targets), parameters (see Tables 4 and 5). These ANN-based mod-
and ’perform’ (performance of the network model). In NARX els are all under the same group, and in fact, the RNN and
model, the function ’closeloop’ was used to replace the feed- CFNN are the older versions of the NARX model. That is
back input with a direct connection from the output layer. The why the study decides to use these models for comparative
hardware specifications including CPU model i5-8250U @ performance analysis. This analysis presents the cogency of
1.60 GHz, GPU model NVIDIA GeForce MX150, and 16 the proposed deep neural network-based approaches in wind
GB RAM, were used for conducting the experiments. time series forecasting.
Though there exists correlation between wind-energy and According to the objectives of the study, three proposed
wind-speed, wind-speed was utilized as true-outputs in train- non-linear models (NARX-NN, NAR-NN, and NIO-NN)
ing the models with external inputs (like temperature and have been investigated to find out the individual performance

Table 4 The best WSP performances of NARX-NN, NAR-NN, and NIO-NN models with multiple inputs
WSP model H&D MSE R2 Epoch Training time

NARX-NN 20H and 1D 1.30786 e−5 9.99860 e−1 245 20 s

NAR-NN 20H and 1D 1.10730 e−4 9.98822 e−1 17 05 s
NIO-NN 20H and 2D 1.32714 e−4 9.98587 e−1 14 10 s

Table 5 The prediction performance comparison of the NARX model with RNN and CFNN models
Models H&D MSE R2 Training time Remarks

NARX 20H & 1D 1.30786e−5 9.99860e−1 20 s Less computation and training time
RNN 10H & 2D 2.6221e−06 9.99970e−1 6.17 min More computation and training time
CFNN 20H & 2D 2.63631e−4 9.99971e−1 2.35 min More computation and training time

123
Journal of Reliable Intelligent Environments (2023) 9:183–200 195

(a)
(a)

(b)
(b)
Fig. 15 WSP results comparisons among NARX-NN, RNN and CFNN
models

NN based model is slightly lower as compared to the other

models (e.g., M S E difference = |1.30786e−5 − 2.6221e−6 |
(c)
= 0.00001 and R 2 difference = |0.999860 − 0.999971| =
Fig. 14 MSE graphs of the NARX-NN, NAR-NN, and NIO-NN models 0.000111), which is negligible in ANN-based time-series
in training phase forecasting with respect to training time (see Table 5 and
Fig. 15).
Moreover, the proposed NARX-NN-based model swiftly
in the wind-speed prediction. The experiments have been showed the prediction outcomes in the case of low training-
done with all external inputs, and Table 4 shows the best time. The prediction errors are evaluated for training, test,
performances of these models in forecasting wind-speed. and validation utilizing an error histogram, as presented in
Figure 14 graphically shows the training performance of Fig. 16. The figure shows the distribution of errors in pre-
NARX-NN, NAR-NN, and NIO-NN-based prediction mod- diction, grouped into 20 bins in each instance. Figure 16a
els in short-term scheme. The prediction outcomes reflected shows that the majority of errors’ distribution is closer to zero
the better achievement of the proposed NARX based model in the NARX-NN model, which represents a fairly spotless
compared to other non-linear models. fit. The auto-correlations are compared through the predic-
Usually, the error decreases after more training, yet it may tion to test the model performance, as illustrated in Fig. 17.
start to rise on the validation dataset as the training may begin Auto-correlations describe prediction errors in association
to over-fit. Commonly the default setting is adjusted to stop with time. In a perfect predicting result, just one non-zero of
the training after six consecutive rises in a validation error. auto-correlation occurs at the zero-lag.
Similarly, the performance of the NARX-NN model was This means that prediction errors are independent on each
compared to other two ANN-based models (such as RNN other, excluding the one at zero lag. The auto-correlation
and CFNN), and the performance measures are shown in helps to develop the training structure that provided bet-
Table 5. Figure 15 illustrates the comparison graphs among ter performance with accurate output for the proposed
NARX-NN, RNN, and CFNN models. models. Based on the experimental results, the proposed
From the experiment results, it was seen that the RNN prediction model using NARX approach has a significant
model performed better in terms of MSE values (2.6221e−6 ), auto-correlation compared to other ANN-based models.
and the NARX-NN and CFNN remained at second and third The input/error given by cross/correlation describes the
place, respectively. In Table 5, the correlation-coefficient correlation of errors with the input pattern; if all correla-
value (R 2 ) belonging to the RNN model is equal to 0.999970 tions are equal to zero, then the prediction model is perfect.
. This R 2 value represents a perfect fit model by predic- Figure 18 presents the correlation of errors for NARX-NN,
tion, and is closer to 1, which means better accuracy. The RNN, and CFNN based prediction models. The experi-
regression values of the CFNN and NARX-NN are 0.999971 ments reflected that the CFNN model has a better input-error
and 0.999860, respectively. The results indicated that the cross-correlation nearby zero-value compared to others. In
performance metrics (M S E and R 2 ) in the proposed NARX- contrast, the NARX-NN model has a fixed input-error cross-

123
196 Journal of Reliable Intelligent Environments (2023) 9:183–200

(a)
(a)

(b)
(b)

(c)
(c) Fig. 17 Auto-correlations of the WSP errors of NARX-NN, RNN, and
CFNN models
Fig. 16 Error-Histogram of the NARX-NN, RNN, and CFNN models

correlation in all bins and the numbers are so close to

zero. Also, the NARX model shows better performance
in learning capability (training time). Thus, the proposed
NARX-NN model performs better in auto-correlations, error-
histogram, and input-error cross-correlation than RNN, (a)
based on Figs. 16, 17 and 18. The NARX-NN model has
less complexity as it is a subclass of the RNN model and
utilizes confined feedbacks from the output layers instead of
hidden layers [94]. Also, the CFNN model is impressible to
the input data, and in the case of new wind-speed conditions,
the curve-fitting model may produce uncertain results [95]
(also shown in Figs. 16, 17 and 18) that the CFNN and RNN
models performs poorly compared to the NARX-NN model. (b)
Analyzing the performance parameters in the experimental
results, it is concluded that the proposed NARX-NN model
performs better prediction in short-term scheme.
Although the proposed approach has been evaluated with
a single data set, it can be applied to similar problems and
additional time-series data set. In addition, we have inves-
tigated another study [96] where the authors evaluated the (c)
NARX model’s performance with some other conventional
methods (such as “feed-forward neural network”, “time- Fig. 18 The input-error cross-correlation of the NARX-NN, RNN, and
delay neural network”, “multilayer perception”, etc.) and CFNN models
similar time-series data set. The study found that the NARX
model performed better compared to other methods with the
MSE value of 1.15940e−1 . In contrast, the MSE value of
the proposed model is equal to 1.30786e−5 in this study.

123
Journal of Reliable Intelligent Environments (2023) 9:183–200 197

The better performance may depend on several issues, such All the prediction models have been developed in Mat-
as the samples of dataset, multiple hidden neurons, delay lab platform, and the experiments used the dataset obtained
time, and the learning algorithms. According to the results from the NREL wind database. This study evaluated the
of this study and the previous findings, it is clear that the performances of the developed models using different mea-
NARX model has a significant prediction performance in surement metrics of MSE and R 2 , auto-correlations, error-
the short-term WSP. This study will help the researchers histogram, the input-error cross-correlation, as well as train-
in better understanding of the deep neural network based ing time. The experimental results revealed that the proposed
models to degrade the computational complexity of the learn- NARX-based model has a better prediction performance
ing process as well as to improve the WSP‘s performance. compared to other models in the terms of auto-correlations,
The literature and experiments showed that many factors error histogram, input-error cross-correlation, and training
could influence the forecasting performance, such as non- time. While the RNN and CFNN models performed better
linearity and uncertainties of time series, variations in input prediction accuracy with R 2 and MSE values, the differences
parameters and datasets, learning algorithms, and model opti- of these values with the NARX model’s performance met-
mizations. In addition, the larger size of the database has rics are minimal (i.e., both are closest to zero). This most
posed new challenges in the data preprocessing and forecast- minor performance difference is negligible in ANN-based
ing process. As input features directly affect the forecasting time-series prediction regarding computational complexity
output, it must focus on an appropriate method for select- and training time. Since the NARX prediction model had
ing local features to balance accuracy and efficiency. An a better efficiency in computation and learning time, this
accurate short-term prediction with NARX network models study decided to give the score to the NARX model. The
can minimize these negative impacts. In addition, this short- proposed approach will be employed to forecast other renew-
term prediction can benefit reliable wind-speed prediction in able energies, like solar-energy and hydropower. As trends
hybrid renewable energy management systems. In practice, a and seasonality are influential characteristics of time series
perfect score does not exist for time series prediction. Hence, data, future works will handle these issues and analyze their
the prediction performance is relative, and most researchers influences in prediction performance with additional time-
investigated previously developed models to compare their series data. Further works will also incorporate a deep neural
newly proposed methods. The MATLAB codes for the devel- network approach with fuzzy logic (i.e., deep neuro-fuzzy
opment of prediction models, results, and data have been model) for estimating energies in hybrid renewable energy
shared as open-source and available in GitHub repository systems.
(https://github.com/mijancse/Deep-Neural-Network-
Acknowledgements The authors would like to acknowledge the sup-
based-Approaches-in-Wind-Time-Series-Forecasting) [97]. port provided by the Institute of Sustainable Energy (ISE) of the Uni-
versiti Tenaga Nasional (@UNITEN, The Energy University, Malaysia)
through BOLD 2025. The authors would like to thank the Department
5 Conclusion of Computer Science and Engineering, Jatiya Kabi Kazi Nazrul Islam
University (@JKKNIU, Bangladesh).

As generating wind-power is subjective to wind-speed, it

Declaration
requires to estimate wind-speed in wind turbines to produce
energy. In hybrid energy systems, the energy forecasting is
a crucial task as it utilizes a renewable energy source. In the Conflict of interest The authors declare that they have no conflict of
case of wind-power generation, precise forecasting of wind- interest.
speed is a complex task because of the uncertain nature of
wind data and its dependency on ambient conditions. The
benefit of utilizing the neural network approaches is lie-down
in its simplicity, and capable of learning nonlinear mod-
els with an input/output relationship. The current prediction References
analysis states that the deep neural network models can pre-
cisely predict the wind-speed as well as power. This paper 1. Shoaib M, Siddiqui I, Rehman S, Khan S, Alhems LM (2019)
Assessment of wind energy potential using wind energy conversion
presents a comprehensive study of the deep neural network system. J Clean Prod 216:346–360
models in wind time-series forecasting. The study aimed to 2. CBS, PBL, RIVM, and WUR (2020) Renewable energy use,
investigate the performance of the proposed NARX-based 1990-2019 (indicator 0385, version 36, 8 september 2020), www.
prediction models, as well as to compare it’s performance environmentaldata.nl. Statistics Netherlands (CBS), The Hague;
PBL Netherlands Environmental Assessment Agency, The Hague;
with other nonlinear models (such as NIO-NN, NAR-NN, RIVM National Institute for Public Health and the Environment,
RNN, and CFNN models) for estimating wind-speed in short- Bilthoven; and Wageningen University and Research, Wageningen.
term basis. Accessed 5 Dec 2020

123
198 Journal of Reliable Intelligent Environments (2023) 9:183–200

3. Sönnichsen, N (2020) Installed wind power capacity—worldwide ing model for load forecasting in smart grid with a comprehensive
2001–2019, Statista: Energy & Environment. https://www. survey. IEEE Access
statista.com/statistics/268363/installed-wind-power-capacity- 23. Jifri MH, Hassan EE, Miswan NH (2017) Forecasting performance
worldwide/. Accessed 5 Dec 2020 of time series and regression in modeling electricity load demand.
4. Sonnichsen N (2020) Global cumulative installed capacity In: 2017 7th IEEE international conference on system engineering
of wind power by country 2019, Statista: Energy & Envi- and technology (ICSET)
ronment. https://www.statista.com/statistics/217522/cumulative- 24. Al-Wadia M, Ismail MT (2011) Selecting wavelet transforms
installed-capacity-of-wind-power-worldwide/. Accessed 5 Dec model in forecasting financial time series data based on arima
2020 model. Appl Math Sci 5(7):315–326
5. Wan C, Lin J, Wang J, Song Y, Dong ZY (2016) Direct quan- 25. Focken U, Lange M, Waldl H-P (2001) Previento-a wind power
tile regression for nonparametric probabilistic forecasting of wind prediction system with an innovative upscaling algorithm. In: 2001
power generation. IEEE Trans Power Syst 32(4):2767–2778 Proceedings of the European wind energy conference, Copen-
6. Zhao Y, Ye L, Li Z, Song X, Lang Y, Su J (2016) A novel bidi- hagen, Denmark, vol 276. Citeseer, conference proceedings
rectional mechanism based on time series model for wind power 26. Nielsen TS, Madsen H, Nielsen HA, Landberg L, Giebel G (2002)
forecasting. Appl Energy 177:793–803 Prediction of regional wind power. In: Proc. of the 2002 global
7. Duran MA, Filik B (2015) Short-term wind speed prediction using windpower conference, conference proceedings
several artificial neural network approaches in eskisehir. In: 2015 27. Giebel G, Landberg L, Nielsen TS, Madsen H (2001) The zephyr
international symposium on innovations in intelligent systems and project-the next generation prediction system. In: Proc. of the 2001
applications (INISTA), IEEE, Conference Proceedings, pp 1–4 European wind energy conference, EWEC’01, Copenhagen, Den-
8. Qiu D, DinçSer H, Yüksel S, Ubay GG (2020) Multi-faceted anal- mark, Conference Proceedings, pp 777–780
ysis of systematic risk-based wind energy investment decisions in 28. Lang S, Mohrlen C, Jorgensen J, Gallachóir B, McKeogh E (2006)
e7 economies using modified hybrid modeling with it2 fuzzy sets. Aggregate forecasting of wind generation on the irish grid using a
Energies 13(6):1423 multi-scheme ensemble prediction system,. In: 2006 International
9. Marsadek M, Mohamed A, Nizam M, Norpiah Z (2008) Risk based solar energy society UK section-conference-c, vol 85, conference
static security assessment in a practical interconnected power sys- proceedings, p 89
tem. In: 2008 IEEE 2nd international power and energy conference, 29. Catalão JPDS, Pousinho HMI, Mendes VMF (2009) An artificial
IEEE, conference proceedings, pp 1619–1622 neural network approach for short-term wind power forecasting in
10. Han S, Qiao Y-H, Yan J, Liu Y-Q, Li L, Wang Z (2019) Mid-to-long portugal. In: 2009 15th international conference on intelligent sys-
term wind and photovoltaic power generation prediction based on tem applications to power systems, IEEE, conference proceedings,
copula function and long short term memory network. Appl Energy pp 1–5
239:181–191 30. Zameer A, Khan A, Javed SG (2015) Machine learning based short
11. Faniband YP, Shaahid S (2000) Forecasting wind speed using arti- term wind power prediction using a hybrid learning model. Comput
ficial neural networks-a case study of a potential location of saudi Eng Electr 45:122–133
arabia. In: E3S Web of Conferences, vol 173. EDP Sciences, con- 31. Alshehri J, Alzahrani A, Khalid M (2019) Wind energy conversion
ference proceedings, p 01004 systems and artificial neural networks: Role and applications. In:
12. Sarma R, Dwight RP, Viré A (2020) Aeroelastic validation and 2019 IEEE innovative smart grid technologies-Asia (ISGT Asia),
bayesian updating of a downwind wind turbine. Wind Energy IEEE, conference proceedings, pp 1777–1782
23(4):864–883 32. Wu G, Liu Y, Wang T (2007) Methods and strategy for modeling
13. Jiang Y, Xingying C, Kun Y, Yingchen L (2017) Short-term wind daily global solar radiation with measured meteorological data-a
power forecasting using hybrid method based on enhanced boost- case study in nanchang station, china. Energy Convers Manage
ing algorithm. J Modern Power Syst Energy Clean 5(1):126–133 48(9):2447–2452
14. Peng H, Liu F, Yang X (2013) A hybrid strategy of short term wind 33. Perdomo R, Banguero E, Gordillo G (2010) Statistical model-
power prediction. Renew Energy 50:590–595 ing for global solar radiation forecasting in bogotá. In: 2010 35th
15. Lei M, Shiyan L, Chuanwen J, Hongling L, Yan Z (2009) A review IEEE photovoltaic specialists conference, IEEE, conference pro-
on the forecasting of wind speed and generated power. Renew Rev ceedings, pp 002 374–002 379
Sustain Energy 13(4):915–920 34. Parras-Gutierrez E, Rivas VM (2010) Time series forecasting: auto-
16. Lange M, Focken U (2008) New developments in wind energy matic determination of lags and radial basis neural networks for a
forecasting. In: 2008 IEEE power and energy society general changing horizon environment. In: The 2010 international joint
meeting-conversion and delivery of electrical energy in the 21st conference on neural networks (IJCNN), IEEE, conference pro-
century, IEEE, Conference Proceedings, pp 1–8 ceedings, pp 1–7
17. Landberg L (1999) Short-term prediction of the power production 35. Hossain MA, Rahman MM, Prodhan UK, Khan MF (2013) Imple-
from wind farms. J Wind Eng Aerodyn Ind 80(1–2):207–220 mentation of back-propagation neural network for isolated bangla
18. Landberg L, Watson SJ (1994) Short-term prediction of local wind speech recognition. Int J Inf Sci Tech 3:4
conditions. Bound-Layer Meteorol 70(1–2):171–195 36. Tascikaraoglu A, Uzunoglu M (2014) A review of combined
19. Rodrigues A, Lopes JP, Miranda P, Palma L, Monteiro C, Bessa R, approaches for prediction of short-term wind speed and power.
Sousa J, Rodrigues C, Matos J (2007) Eprev-a wind power fore- Renew Rev Sustain Energy 34:243–254
casting tool for portugal. In: 2007 Proceedings of the European 37. Amjady N, Keynia F, Zareipour H (2011) Wind power prediction by
wind energy conference EWEC, vol 7, conference proceedings a new forecast engine composed of modified hybrid neural network
20. Thordarson F, Madsen H, Nielsen HA, Pinson P (2010) Condi- and enhanced particle swarm optimization. IEEE Trans Sustain
tional weighted combination of wind power forecasts. Wind Energy Energy 2(3):265–276
13(8):751–763 38. Jung J, Broadwater RP (2014) Current status and future advances
21. Svensson M (2015) Short-term wind power forecasting using artifi- for wind speed and power forecasting. Renew Rev Sustain Energy
cial neural networks, Master’s Thesis, School of Computer Science 31:762–777
and Communication, Royal Institute of Technology KTH 39. Lawan S, Abidin W, Chai W, Baharun A, Masri T (2014) Different
22. Akhtaruzzaman M, Hasan MK, Kabir SR, Abdullah SNHS, Sadeq models of wind speed prediction; a comprehensive review. Int J Sci
MJ, Hossain E (2020) Hsic bottleneck based distributed deep learn- Res Eng 5(1):1760–1768

123
Journal of Reliable Intelligent Environments (2023) 9:183–200 199

40. Kariniotakis G, Stavrakakis G, Nogaret E (1996) Wind power 60. Cadenas E, Rivera W, Campos-Amezcua R, Heard C (2016) Wind
forecasting using advanced neural networks models. IEEE Trans speed prediction using a univariate arima model and a multivariate
Energy Convers 11(4):762–767 narx model. Energies 9(2):109
41. Riahy GH, Abedi M (2008) Short term wind speed forecasting for 61. Zhang J, Yin Z, Wang R (2017) Nonlinear dynamic classification of
wind turbine applications using linear prediction method. Renew momentary mental workload using physiological features and narx-
Energy 33(1):35–41 model-based least-squares support vector machines. IEEE Trans
42. Fadare D (2010) The application of artificial neural networks to Hum-Mach Syst 47(4):536–549
mapping of wind speed profile for energy application in nigeria. 62. Nedjah N, de Macedo Mourelle L (2016) Online pattern recognition
Appl Energy 87(3):934–942 for portuguese phonemes using multi-layer perceptron combined
43. Rahman MM, Bhuiyan A-A (2015) Comparison study and result with recurrent non-linear autoregressive neural networks with
analysis of improved back-propagation algorithms in bangla speech exogenous inputs. In: 2016 IEEE Latin American conference on
recognition. Int J Appl Res Inf Technol Comput 6(2):107–117 Computational Intelligence (LA-CCI), IEEE, Conference Proceed-
44. Filik B, Filik T (2017) Wind speed prediction using artificial neu- ings, pp 1–6
ral networks based on multiple local measurements in eskisehir. 63. Braga DD, Tanscheit R, Vellasco MM (2017) Neural network
Energy Procedia 107:264–269 nonlinear plant identification as a tool in intelligent controller
45. Kusiak A, Zhang Z (2010) Short-horizon prediction of wind power: design. In: 2017 International Joint Conference on Neural Net-
A data-driven approach. IEEE Trans Energy Convers 25(4):1112– works (IJCNN), IEEE, conference proceedings, pp 1472–1479
1122 64. Hoyle N, Flow E (2009) Wind turbine power calculations. RWE n
46. Catalão JPDS, Pousinho HMI, Mendes VMF (2011) Short-term Power Renew 1(1):1–5
wind power forecasting in portugal by neural networks and wavelet 65. Santhosh M, Venkaiah C, Vinod-Kumar D (2020) Current advances
transform. Renew Energy 36(4):1245–1251 and approaches in wind speed and wind power forecasting for
47. Jursa R, Rohrig K (2008) Short-term wind power forecasting using improved renewable energy integration: A review. J Eng Rep
evolutionary algorithms for the automated specification of artificial 2020:e12178
intelligence models. Int J Forecast 24(4):694–709 66. Engineering-ToolBox (2009) Wind power. https://www.
48. Cao Q, Ewing BT, Thompson MA (2012) Forecasting wind speed engineeringtoolbox.com/wind-power-d_1214.html. Accessed
with recurrent neural networks. Eur J Oper Res 221(1):148–154 6 Dec 2020
49. Liangyou H, Dongxiang J, Qian H, Yongshan D (2002) Wind speed 67. Brown BG, Katz RW, Murphy AH (1984) Time series models to
forecasting using fully recurrent neural network in wind power simulate and forecast wind speed and wind power. J Clim Appl
plants. IEEE Trans Energy Convers 11(762–767):32 Meteorol 23(8):1184–1195
50. Qin Q, Lai X, Zou J (2019) Direct multistep wind speed forecasting 68. Sohoni V, Gupta S, Nema R (2016) A critical review on wind
using lstm neural network combining eemd and fuzzy entropy. Appl turbine power curve modelling techniques and their applications in
Sci 9(1):126 wind based energy systems. J Energy 2016:5
51. Liang S, Nguyen L, Jin F (2018) A multi-variable stacked long- 69. Okumus I, Dinler A (2016) Current status of wind energy forecast-
short term memory network for wind speed forecasting. In: 2018 ing and a hybrid method for hourly predictions. Energy Convers
IEEE international conference on big data (Big Data), IEEE, con- Manage 123:362–371
ference proceedings, pp 4561–4564 70. Shetty RP, Sathyabhama A, Pai PS (2020) Comparison of modeling
52. Zhang W-Y, Zhao Z-B, Han T-T, Kong L-B (2011) Short term wind methods for wind power prediction: a critical study. Front Energy
speed forecasting for wind farms using an improved autoregression 14(2):347–358
method. In: 2011 international conference of information technol- 71. Wang X, Guo P, Huang X (2011) A review of wind power fore-
ogy, computer engineering and management sciences, vol 1, IEEE, casting models. Energy Procedia 12:770–778
conference proceedings, pp 195–198 72. Costa A, Crespo A, Navarro J, Lizcano G, Madsen H, Feitosa E
53. Nazaripouya H, Wang B, Wang Y, Chu P, Pota H, Gadh R (2016) (2008) A review on the young history of the wind power short-term
Univariate time series prediction of solar power using a hybrid prediction. Renew Rev Sustain Energy 12(6):1725–1744
wavelet-arma-narx prediction method. In: 2016 IEEE/PES Trans- 73. Kusiak A, Zheng H, Song Z (2009) Short-term prediction of wind
mission and Distribution Conference and Exposition (T&D), IEEE, farm power: a data mining approach. IEEE Trans Energy Convers
conference proceedings, pp 1–5 24(1):125–136
54. Pasari S, Shah A, Sirpurkar U (2020) Wind energy prediction using 74. Treiber NA, Späth S, Heinermann J, von Bremen L, Kramer O
artificial neural networks. Springer, Cham, pp 101–107 (2020) Comparison of numerical models and statistical learning for
55. Di Piazza A, Di Piazza MC, Vitale G (2016) Solar and wind fore- wind speed prediction. In: Proceedings of the European symposium
casting by narx neural networks. Renew Energy Sustain Environ on artificial neural networks, ESANN, conference proceedings, pp
1:39 71–76
56. Araüjo RDA, Oliveira AL, Meira S (2017) A morphological neural 75. Xiaofeng L, Chunshan Y (2014) Research and application of data
network for binary classification problems. Eng Appl Artif Intell mining and narx neural networks in load forecasting. Int J Database
65:12–28 Theory Appl 7(2):13–24
57. Boussaada Z, Curea O, Remaci A, Camblong H, Mrabet-Bellaaj 76. Xiao L, Zhang J, Zhu T, Shi C, Zhang H (2007) Analysis of urban
N (2018) A nonlinear autoregressive exogenous (narx) neural net- power load based on association rules [j]. Automation of electric
work model for the prediction of the daily direct solar radiation. power systems, vol 17
Energies 11(3):620 77. Li Y-H, Lei J-H (2011) The application and research of electric
58. Chen S, Billings S, Grant P (1990) Non-linear system identification power load forecasting technology based on the time series model.
using neural networks. Int J Control 51(6):1191–1214 Sci Technol Eng 11(4):860–864
59. Li H, Zhu Y, Hu J, Li Z (2017) A localized narx neural network 78. Rahman MM, Shakeri M, Tiong SK, Khatun F, Amin N, Pasupuleti
model for short-term load forecasting based upon self-organizing J, Hasan MK (2021) Prospective methodologies in hybrid renew-
mapping. In: 2017 IEEE 3rd International Future Energy Electron- able energy systems for energy prediction using artificial neural
ics Conference and ECCE Asia (IFEEC 2017-ECCE Asia), IEEE, networks. Sustainability 13(4):2393
Conference Proceedings, pp 749–754

123
200 Journal of Reliable Intelligent Environments (2023) 9:183–200

79. Lin T, Horne BG, Tino P, Giles CL (1996) Learning long-term 91. Hagan MT, De Jesús O, Schultz R, Medsker L, Jain L (1999)
dependencies in narx recurrent neural networks. IEEE Trans Neural Training recurrent networks for filtering and control. CRC Press,
Netw 7(6):1329–1338 Hoboken, pp 311–340
80. Abdulkadir SJ, Yong S-P (2015) Scaled ukf-narx hybrid model 92. Mandic D, Chambers J (2001) Recurrent neural networks for pre-
for multi-step-ahead forecasting of chaotic time series data. Soft diction: learning algorithms, architectures and stability. Wiley,
Comput 19(12):3479–3496 Hoboken
81. Siegelmann HT, Horne BG, Giles CL (1997) Computational capa- 93. Younis O (2018) A predictive model for solar photovoltaic power
bilities of recurrent narx neural networks. IEEE Trans Syst Man using the levenberg-marquardt and bayesian regularization algo-
Cybern Part B 27(2):208–215 rithms and real-time weather data. Int J Adv Comput Sci Appl 9:1
82. Xie H, Tang H, Liao Y-H (2009) Time series prediction based 94. Lipu MSH, Hannan MA, Hussain A, Saad MHM, Ayob A, Blaab-
on narx neural networks: An advanced approach. In: 2009 Inter- jerg F (2018) State of charge estimation for lithium-ion battery
national conference on machine learning and cybernetics, vol 3, using recurrent narx neural network model based lighting search
IEEE, conference proceedings, pp 1275–1279 algorithm. IEEE Access 6:28-150-28–161
83. Diaconescu E (2008) The use of narx neural networks to predict 95. Safiullin N, Porshnev S, Kleeorin N (2017) Forecast of monthly
chaotic time series. Wseas Trans Comput Res 3(3):182–191 sunspot numbers using nonlinear dynamo model with neural net-
84. Li G, Shi J (2010) On comparing three artificial neural networks works. In: 2017 Dynamics of systems, mechanisms and machines
for wind speed forecasting. Appl Energy 87(7):2313–2320 (dynamics), conference proceedings, pp 1–4
85. Ferreira AA, Ludermir TB, de Aquino RR (2012) Comparing 96. Noman F, Alkawsi G, Alkahtani AA, Al-Shetwi AQ, Tiong
recurrent networks for time-series forecasting. In: The 2012 inter- SK, Alalwan N, Ekanayake J, Alzahrani AI (2020) Multistep
national joint conference on neural networks (IJCNN), IEEE, short-term wind speed prediction using nonlinear auto-regressive
conference proceedings, pp 1–8 neural network with exogenous variable selection. Alex Eng J
86. Fan S, Liao JR, Yokoyama R, Chen L, Lee W-J (2009) Forecasting 60(1):1221–1229
the wind generation using a two-stage network based on meteoro- 97. Rahman MM (2021) Deep neural network based approaches in
logical information. IEEE Trans Energy Convers 24(2):474–482 wind time series forecasting. Github Repository. https://github.
87. Brower M (2020) Development of eastern regional wind com/mijancse/Deep-Neural-Network-based-Approaches-in-
resource and wind plant output datasets: March 3, 2008–march Wind-Time-Series-Forecasting. Accessed 22 Sept 2021
31, 2010, National Renewable Energy Lab. (NREL), Golden,
CO (United States). https://www.nrel.gov/grid/eastern-wind-data.
html. Accessed 12 Oct 2020
Publisher’s Note Springer Nature remains neutral with regard to juris-
88. Ramirez-Rosado IJ, Fernandez-Jimenez LA, Monteiro C, Sousa J,
dictional claims in published maps and institutional affiliations.
Bessa R (2009) Comparison of two new short-term wind-power
forecasting systems. Renew Energy 34(7):1848–1854
89. Peng H, Long F, Ding C (2005) Feature selection based on mutual
information criteria of max-dependency, max-relevance, and min-
redundancy. IEEE Trans Pattern Anal Intell Mach 27(8):1226–
1238
90. Methaprayoon K, Yingvivatanapong C, Lee W-J, Liao JR (2007)
An integration of ann wind power estimation into unit commit-
ment considering the forecasting uncertainty. IEEE Trans Ind Appl
43(6):1441-1448

123