This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
1

Deep Spatial-Temporal 2-D CNN-BLSTM Model

for Ultra-Short-Term LiDAR-Assisted Wind
Turbine’s Power and Fatigue load Forecasting
Amirhossein Dolatabadi, Graduate Student Member, IEEE, Hussein Abdeltawab, Member, IEEE,, and Yasser
Abdel-Rady I. Mohamed, Fellow, IEEE

Abstract—Optimizing wind turbine performance is still a τm Mechanical time constant of the wind turbine.
challenge due to the dynamic interactions between the spatially- τe Electrical time constant of the wind turbine.
temporally stochastic wind fields and the wind turbine as a ρ Air density.
complex mechanical system. Recent cost reduction of remote
sensing wind measurement technologies, such as light detection A Swept area of the rotor blades.
and ranging (LiDAR), has opened a new research area on the use R Radius of the rotor blades.
of deep learning models for predicting wind turbine’s responses. Cp Power coefficient of the rotor.
In this paper, a LiDAR-aided deep learning model is presented λ Tip speed ratio.
to learn the powerful spatial-temporal characteristics from the β Pitch angle.
input wind fields. In the proposed method, the combination
of 2-D convolutional neural networks (CNNs) and bidirectional Ct Thrust coefficient.
long short-term memory (BLSTM) units is used to capture N Total number of samples.
high levels of abstractions in wind fields concurrently, and
thus forecasting wind output power and fatigue load as two
representatives of wind turbine responses. The LiDAR wind C. Variables:
preview information is used as the 2-D-images of wind fields
for the CNN. Moreover, the BLSTM is incorporated with the υilos line-of-sight wind speed.
proposed CNN to improve the forecasting accuracy further and u Wind field component vector in x direction.
learn deep temporal features. The aero-elastic 5-MW reference v Wind field component vector in x direction.
wind turbine of National Renewable Energy Laboratory (NREL) w Wind field component vector in x direction.
is used to evaluate the performance of proposed model compared Powt Wind turbine output electrical power.
to the state-of-the-art deep-learning based architectures in the
recent literature. Piwt Wind turbine input mechanical power.
kl Trainable kernel weight vector of the lth layer.
Index Terms—Deep learning, convolutional neural network,
bl Trainable kernel bias vector of the lth layer.
bidirectional long short-term memory, spatial-temporal features,
fatigue load, light detection and ranging (LiDAR). ξl Trainable weight matrix of the lth layer.
f Activation vectors of the forget gate.
z Activation vectors of the input gate.
N OMENCLATURE o Activation vectors of the output gate.
A. Indices: hft Forward hidden states in BLSTM network
t Index of the time sample. hbt Backward hidden states in BLSTM network
i Index of laser beam vector focus point. yb Desired output vector.
l Index of hidden layer for CNN. y Actual output vector.
n Index of output layer sample.
I. I NTRODUCTION
B. Parameters:
a
ϕ
Distance of wind profile from the hub.
Vertical dimension of the measurement grid.
W ITH the boost of concerns about the greenhouse effect
and climate change, integration of renewable energy
sources (RESs) to the energy systems has gained much at-
γ Horizontal dimension of the measurement grid. tention around the world. In particular, the government of
η wt Overall efficiency of energy conversion system. Alberta, Canada has set a firm target for the Alberta electric
This work was supported in part by the Future Energy Systems Research system operator (AESO) to move towards having 30% of
through the Canada First Research Excellence Fund (CFREF) and in part by Alberta’s electricity coming from RESs by the year 2030, with
the Natural Sciences and Engineering Research Council of Canada (NSERC). an estimated additional RESs integration of 5,000 MW to the
(Corresponding author: Amirhossein Dolatabadi.)
Amirhossein Dolatabadi and Yasser Abdel-Rady I. Mohamed are with the grid. It is anticipated that a significant portion of the 5,000
Department of Electrical and Computer Engineering, University of Alberta, MW of renewable energy capacity will come from wind power
Edmonton, AB T6G 2V4, Canada (e-mail: adolatab@ualberta.ca; yasser2@ [1]. While being non-polluting and widely available, wind
ualberta.ca).
Hussein Abdeltawab is with the School of Engineering - Penn State power also could cause considerable challenges to the stability
Behrend, Erie, PA 16563, USA (e-mail: hza5222@psu.edu). and security of the energy system due to its stochasticity,

1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
uncertainty, and intermittency [2], [3]. Thus, in this chang-
ing environment, an accurate and stable prediction of wind
turbine responses is a pressing issue for operational planning,
scheduling and real-time balancing of large-scale wind power
integration.
Fatigue loads induced by sudden variations of the wind can
increase maintenance cost and reduce the operating life of
wind turbine. In this regard, various yaw misalignment and
de-loading control methods have been presented to increase
power generation efficiency and mitigate the mechanical loads
on the wind turbine structure [4]. From the wind farms control
engineering point of view, low maintenance cost and efficient
energy harvesting in wind turbines can be achieved by more
precise and superior wind fields measurements methodologies
instead of using traditional technologies, such as the anemome-
ter and nacelle vane, in which only close wind flows in the
vicinity of sensor location can be measured [5].
Broadly speaking, according to the length of the forecasting
horizon, the wind turbine responses prediction techniques can
be mainly categorized into: 1) ultra-short-term and short-term,
2) medium-term, 3) long-term ones [6]. An ultra-short-term
and short-term forecasting methods refer to the prediction
of wind turbine dynamic responses in the range of a few
seconds to few minutes and few minutes to hours ahead,
respectively. These tasks are mainly applied for real-time
operations and electricity market clearing [7]. A medium-term
prediction model seeks to forecast wind turbine output data
at time intervals of several hours to one week. This type of
prediction is generally benefiting unit commitment decisions
and energy storage system (ESS) scheduling [8]. The Long-
term forecasting model focuses on the time scale of one week
to a year or more to optimize expansion and maintenance
planning [9].
In the technical literature, wind turbine responses forecast-
ing is generally made using a wind speed time series measured
at the hub height of the turbine. This highly varying time series
has a chaotic and stochastic characteristic; thus, forecasting
can be considered as a complex regression task. With a pack
of meteorological factors and boundary conditions, a physical
forecasting model was built in [10]. Some forecasting models
are statistical-based, which try to capture mathematical rela-
tionships in the historical data. For example, an autoregressive
integrated moving average (ARIMA) approach was used to
represent the upper and lower bounds of the wind power
generation in [11]. However, the linear nature of statistical
models restricts their ability to handle the nonlinear patterns
and deal with challenging wind data prediction problems. On
the other hand, wind turbine responses can be forecasted by
pattern recognition and machine learning techniques such as
support vector regression (SVR) and artificial neural networks
(ANNs). Using wind speed data, a wind power prediction
model based on feed-forward neural network (FFNN) was
proposed in [12]. Recurrent neural networks (RNNs) which
forecast the future value by using the current inputs and the
experience have also been investigated to improve the forecast
performance [13].
Conventional shallow ANNs forecast models as mentioned
above suffer from the problem that they cannot efficiently learn
1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
2
sophisticated features, and thus the prediction accuracy may
be highly variable. With the booming advancement of machine
learning techniques, recent research of wind turbine responses
prediction has focused more on deep models which can effec-
tively extract inherent abstract features of the highly varying
time series data. Commonly seen networks include combina-
tion of interval probability distribution learning (IPDL) and
deep belief network (DBN), gated recurrent unit (GRU) and
long short-term memory (LSTM). Especially, LSTM architec-
ture, which was initially proposed by Hochreiter and Schmid-
huber [14], opened a new research area in the realm of deep
networks. In very recent literature, a hybrid forecasting model
based on stacked denoising autoencoders and LSTM network
was developed to predict wind speed data in [15]. As another
example, a two-layer nonlinear combination method based
on extreme learning machine (ELM), Elman neural network
(ENN), and LSTM for short-term wind speed prediction was
introduced in [16].
Nevertheless, the aforementioned approaches focus on the
wind speed time series measured at the hub heights of turbine,
which ignore the chaotic and stochastic characteristics of wind
turbulent flow over a rotor area. Thus, there is a pressing
need for considering more extensive feature measurements
of inflowing wind towards the turbine, such as the speed,
direction and turbulence. Recently, as an advanced remote
sensing wind measurement technology, the light detection and
ranging (LiDAR) has been proposed and attracted extensive
attention to improve inaccurate and unstable measurements of
nacelle vane [17]. Moreover, besides higher accuracy, other ad-
vantages of the LiDAR over the conventional mechanical wind
vanes and anemometers used in wind turbines are measuring
diverse and longer distances and also flexible installation. For
example, in [18], the annual energy production increased by
1.83% through a LiDAR-based method for yaw error align-
ment. Results from [19] showed that LiDAR-aided wind speed
measurement yielded fatigue or extreme load reduction on the
wind turbine tower by 10%, thus improving the operation life
of wind turbine. Comprehensive analyses and discussions on
the LiDAR-based wind turbine performance can be found in
[20]. Although different efforts have been made to improve
LiDAR-assisted control of wind turbines, minimal studies
are made on integrating such extensive wind field data and
machine learning approaches to find wind turbine responses. In
[21], a FFNN was employed for the extrapolation of the higher
heights wind speed using lower heights values. However,
the LiDAR measurements were only based on single hub
height values, and thus, did not consider the dynamics of
the turbulent wind field. Table I summarizes a taxonomy of
recently proposed models in LSTM and CNN combination
area. To predict wind turbine’s responses with satisfactory
accuracy, three main challenges have to be addressed. (1) Most
data-driven methods have employed 1-D and 2-D networks
in a serial manner. This assumption creates some error in
the forecasted results as the extracted features from the first
network have a significant influence on the training of the
next network. (2) The proposed one-directional LSTM models,
e.g., [31], introduce some error in spike points forecasting as
they solely obey the recursive procedure which fed back the
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
3

previous information in an iterative manner. (3) Most CNN-
based approaches have just converted 1-D time series data to a
matrix form and then apply it to the CNN instead of using 2-
D wind field images, e.g., [32]–[35]. These time series-based
matrices cannot model the complex interactions between the
turbine and a sequence of wind fields.
To bridge this gap, we seek to address two important issues
for wind power operators. Is it possible to apply deep learning
to the data of a LiDAR-enhanced wind turbine? Can image-
based deep learning approaches really increase the accuracy
of forecasting compared with other techniques? To this end,
this paper proposes a novel LiDAR-assisted deep neural net-
work model to better forecast ultra-short-term wind turbine
response. Complex wind fields measured by the LiDAR are
employed as a sequence of images for the deep convolutional
neural network (CNN) and time series of different wind speed
components are handled by the bidirectional LSTM (BLSTM)
network to learn deep spatial and temporal features of tur-
bulent wind flow simultaneously. Considering the limitations
of previous models and research works, we propose a novel
data-driven LiDAR-assisted framework, namely 2-D CNN-
BLSTM, for predicting wind turbine’s responses and our main
contributions summarized as follows.
To the best of our knowledge, this is the first effort
to incorporate the 2-D CNN with recurrent BLSTM to
concurrently capture the spatial and temporal features of
a turbulent wind field.
Wind profiles in front of the turbine’s blades are used to
predict future wind turbine responses instead of using a
single hub-height wind speed time series.
The proposed 2-D CNN-BLSTM framework can better
handle the uncertainties and learn deep temporal features
from the sequential input data using recurrent BLSTM
units, which utilizes the previous and future hidden layer
features, compared to single-network models.
The proposed LiDAR-assisted deep model provides ultra-
short-term future measurements according to upcoming
wind flows before reaching the turbine blades, rather than
considering those which have already interacted with the
turbine.
The proposed model provides accurate ultra-short-term
forecasting of fatigue or extreme loads yields more effi-
cient yaw misalignment and de-loading control strategies
for a LiDAR-enhanced wind turbine and consequently
better lifetime and energy harvesting efficiency.
The rest of this paper is structured as follows. Section II
introduces the LiDAR and wind turbine modeling methodolo-
gies. In section III, the overall body of the proposed frame-
work, which consists of 2-D CNN and BLSTM networks, is
presented. The numerical results are presented and discussed
in Section IV. Finally, the conclusion is drawn in Section V.
II. P RINCIPLE OF L I DAR MEASUREMENTS
Generally, there are two types of LiDAR installations: 1)
Nacelle-mounted LiDAR, which similar to the nacelle vane
requires to be deployed on the nacelle rooftop and can measure
freestream wind at 50 to 200 meters distance in front of the
turbine blades as shown in Fig. 1(a). 2) Ground-based LiDAR,
which demands to be located on the ground and can vertically
emit a laser beam to measure freestream wind as shown in
Fig. 1(b). A LiDAR sensor mounted on nacelle can overcome
the above-stated disadvantages, as it is able to provide a
sufficiently early preview measurement of the undisturbed
inflow over the entire rotor area at a far distance. Moreover,
it is worth noting that a nacelle-mounted LiDAR can reliably
determine the speed and direction of the wind regardless of
the rotor turbulence, which is possible by manipulation the
location of the beam in front of the rotor. The wind speed
measurement along the direction of i-th LiDAR laser beam
can be modeled by equations (1) and (2) [36].
Z +∞
υilos = (lix u (a) + liy v (a) + liz w (a))fL (a) da (1)
−∞
2
e−4 ln 2(a/W )
fL (a) = R +∞ 2
(2)
−∞
e−4 ln 2(a/W ) da
where fL (a) is the weighting function at the distance a. The
simplified version of equation (1) can be expressed by equation
(3).
υilos = lix ui + liy vi + liz wi (3)

TABLE I: Comparison of recently published CNN+LSTM studies with the proposed approach..

Reference Forecasting area CNN input Bidirectional process Combination Ensemble learning
[22] Air pollutant 2D time series × Serial ×
[23] Turbofan RUL 2D time series × Serial ×
[24] PMU data 2D time series × Serial ×
[25] Battery RUL 2D time series × Serial X
[26] Battery SOC 1D time series × Serial ×
[27] Load 2D time series × Serial ×
[28] Load 2D time series X Serial ×
[29] PV 1D time series × Serial ×
[30] PV 2D time series × Parallel X
[31] Wind 2D time series × Serial ×
Proposed Wind 2D LiDAR images X Serial X

1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
(a)
Fig. 1: Different types of LiDAR based on installation loca-
tion.(a) Nacelle-mounted LiDAR. (b) Ground-based LiDAR
T 1
where υilos is wind vector [ui vi wi ] projection in the i-th
focus point on the direction of normalized laser beam vector
with the length fi :
 x   
li xi
 liy  = 1  yi  (4)
fi
liz zi
III. BACKGROUND T HEORIES
To enhance the forecasting accuracy and reliability of the
wind turbine response, it is essential to learn the spatial-
temporal patterns, which improves the knowledge of a tur-
bulent wind field. In this section, the structure of the designed
wind turbine response forecasting method is described in full
details.
A. Wind turbine modeling
In real-world, the horizontal axis wind turbine of MW class
typically adopts different active de-loading and nacelle direc-
tion control strategies by using the predicted wind speed and
wind direction data. However, traditional wind field measure-
ment methods such as anemometer and vane have a common
drawback. They can only cover a limited close vicinity of
their location; therefore, these approaches would not have an
appropriate performance to be applied for the wind turbine
response prediction task. On the other hand, accurate wind
data measurement has a crucial role in the optimal harvesting
of wind power and increasing turbine fatigue lifespan by
decreasing the induced parasitic load on the blades and main
bearing. For example, reference [37] showed that for 7.5° to
15° of yaw error, the power loss in a wind turbine lied in the
range from 2.4% to 13%.
The wind speed at height py and distance px from the hub
center of the turbine is υpx , py (ai ), which is measured by
LiDAR at distance ai in front of the turbine. On the other
hand, the wind field at distance ai will reach the turbine after
time ti = υaavg
i
, where υavg denotes the average wind speed.
By assuming uniform average wind speed for the proposed
measurement grid:
υpx , py (ai ) = υpx , py (υavg . ti ) ' υpx , py (t + ti )
2ϕ
px ∈ [−ϕ, ϕ] and py ∈ [−γ, γ] ( ∆p x
= Rpx and
2γ
∆py = Rpy are the resolutions in the vertical and horizontal
directions). The effective uniform speed that will have the
1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
Transactions on Industrial Informatics
4
same effect of wind speed grid at time ti can be defined as
follows:


υ (t) = θ υpx , py (t) ∀ px ∈ [−ϕ, ϕ] and py ∈ [−γ, γ] (6)
The wind turbine output power can be described by the
following equation:
(b)
η wt Piwt (s) η wt P wt (s)
Powt (s) = (τm s+1)(τe s+1) ' τm s+1
i
(7)
= G (s) Piwt (s)
Piwt is the input power of turbine and can be calculated as
follows.
Piwt = ρAR υ 3 Cp (λ, β) (8)
2
υ is the magnitude of the wind velocity component which is
perpendicular to the rotor plane. From equation (8), it can be
seen that wind turbine extracted power increases with the cube
of the υ. Apart from affecting the generated power, inaccurate
estimations of wind speed and wind direction will also have
a significant effect on the turbine’s structural loads. From the
wind farms operator point of view, a premature fatigue failure
problem can be problematic as the drivetrain of a wind turbine
needs much more maintenance cost and downtime than the
other subassemblies of the wind turbine.
Under this background, the most significant issue is how
to calculate the wind turbine’s fatigue load. In this paper, we
consider the tower bending moment as a measure of fatigue
load [4]. The thrust force, which is the main cause of tower
bending moment, is described by the following equation:
1
wind
ρAR υ 2 Ct (λ, β)
Fthrust = (9)
2
The details of the Ct and Cp are stated in [38] as 2-D lookup
tables.
By using the convolution operation, the wind turbine output
power is expressed as:
t+t
R i
Powt (t + ti ) = g (t)Piwt (t + ti − τ ) dτ
0 (10)
= g (t) ∗ Piwt (τ + ti )
Finally, it is worth noting that all the above-mentioned
procedure can be used for wind turbine fatigue load by
replacing Eq. (8) with Eq. (9). The structure of the proposed
deep spatial-temporal feature learning model, namely 2-D
CNN-BLSTM, is visualized in Fig. 2. As presented in the
figure, 2-D wind field images and 1-D wind time series data
are fed to the 2D-CNN and BLSTM deep learning architecture
to extract spatial and temporal features from each wind data,
respectively. These branches operate independently of each
other until they are concatenated.
B. Convolutional neural network (CNN)
Due to the superior ability of conventional algorithms in
(5) solving complex tasks, CNNs have been successfully used
in many areas. CNNs have powerful self-tuning & learning
capability; thus, they can efficiently capture complex spatial
features from the highly varying wind flows. The convolutional
layers provide translation invariance ability by using a set
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
5

...
...

power or fatigue load)

Fully connected layer

Output layer (wind

2-D wind field images Convolution pooling

Concatenate layer
Convolution pooling Flattrning

... LSTM
... LSTM

... LSTM
... LSTM
... LSTM
... LSTM

Fully connected layer

Fully connected layer
...
LSTM
LSTM

LSTM
LSTM
LSTM
LSTM
LSTM
LSTM

LSTM
LSTM
LSTM
LSTM
1-D hub-height wind speed time series BLST M layer BLST M layer BLST M layer

Fig. 2: Overall Structure of 2-D CNN-BLSTM framework.

ht
ht ht+1
Ct-1
ˣ + Ct

ˣ tanh
Ct-1 Ct Ct+1 ...
LSTM LSTM
ft it gt Ot
ˣ block block
ht-1 ht ht+1 ...
σ σ tanh σ
ht-1 ht
xt xt+1
xt

Fig. 3: Schematic of a typical LSTM block. Fig. 4: Illustration of an un-rolled LSTM network.

of learnable kernels and inductive bias of local connectivity, In a deep CNN with multiple filters with different kernel
decreasing the number of learning parameters and, conse- sizes, deep spatial features can be effectively extracted. It
quently, increases generalization capability. Such a convolution is worth noting that this benefit emerged due to the shared
operator is equivalent to moving kernels over spatial positions, weights, local connectivity, and receptive fields features of
which for the q th feature map of the lth layer is defined as CNNs.
given in equation (11) [39].
 
X C. BLSTM Network
xql = f  xpl−1 ⊗ klpq + bql 


(11)
p∈N q RNN has gradually gained attentions as a powerful
sequence-based architecture to process dynamic characteristics
f (.) is a nonlinear activation function, and ⊗ denotes a
of input data along with the development of data-driven
convolution operation.
techniques. In contrast to FFNNs, such models are capable
Then, a pooling layer is integrated to reduce the compu-
of using the data of the previous hidden or output layers with
tational cost of the deep CNN network and consequently the
their internal state (memory) to find the temporal correlations
possibility of over-fitting by down-sampling feature maps. The
between the current information and past circumstances.
pooling operation can be defined mathematically by:
xql = fdown xql−1


(12)
Output
where fdown (.) represents the down-sampling function. ... yt-1 yt yt+
1
... layer

Average-pooling and max-pooling operations are the two Backward

ht-1 ht ht+1
popular pooling strategies in the CNNs. In this work, max- process

pooling is chosen to produce a feature map containing the ht-1 ht ht+1

Forward
process
most prominent features. Finally, the fully-connected layer is
utilized to flatten the 2-D feature maps and make them suitable ... xt-1 xt xt+1 ...
Input
layer

for putting through the activation function as follows:

Fig. 5: General structure of BLSTM framework.
xl = gf c (ξl xl−1 + bl ) (13)

1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
(a)
Fig. 6: Different wind field components generated by TurbSim. (a) streamwise (longitudinal) component u. (b) transverse
(crosswise) component v. (c) vertical component w.
Although RNNs are considered a significant improvement
compared to traditional feed-forward networks, they can suffer
from the exponentially fast decreasing of training backprop-
agated errors in the steepest descent algorithm called gra-
dient vanishing problem. This issue restricts the capability
of the network to learn temporal correlations. To address
this, LSTM architecture as an elaborated version of RNN
with rich dynamics is introduced [14]. A general structure
of the LSTM block is shown in Fig. 3. Each LSTM block
has three main multiplicative units called input, output, and
forget gates for continuous writing, reading, and resetting data,
respectively. Specifically, the new and previous information
are memorized by the input gate, and inconsequential and
irrelevant information is scraped from the memory cell by the
forget gate. The output gate exploits the beneficial information
and controls the impact of memory content on the output.
Different LSTM blocks can be stacked together to compose a
deep LSTM network and propagate the features among these
blocks during the network’s training process. In this situation,
the stochastic trend of complex and challenging temporal
phenomena is captured more effectively. Succinctly, Fig. 4
illustrates the sequential schematic of an un-rolled LSTM
network.
Overall, the computational procedure of the forward pass
in a LSTM architecture can be formulated by equations (14)-
(19):
 
ft = σ Wxf xt + Whf ht−1 + bf (14)
zt = σ (Wxz xt + Whz ht−1 + bz ) (15)
gt = tanh (Wxg xt + Whg ht−1 + bg ) (16)
ot = σ (Wxo xt + Who ht−1 + bo ) (17)
6
(b) (c)
matrices and the bias vector which will be tuned during by
optimization algorithm the training procedure.
Another important issue is that by processing the data
in temporal way, the future context’s information can be
neglected. In this study, to overcome this shortcoming of
the LSTM networks, the bidirectional structure is incorpo-
rated into the proposed LSTM network to handle the whole
temporal horizon’s information. In this way, not only the
previous features are used but also upcoming information is
implemented. The overall structure of the BLSTM concept is
illustrated in Fig. 5. Two different hidden layers are utilized
to bring bidirectional memory; one as a forward hidden layer
for passing information from past to future and the other as a
backward hidden layer for passing information from future to
past. More specifically, when deep learning-based architectures
are built, one can achieve much higher data representation
capability than typical LSTMs. This process prevents error
accumulation and leads to having a more accurate prediction
of highly intermittent and stochastic phenomena. In BLSTM,
the output is explained according to the following equations:
 
hft = tanh xt Wxhf
+ hft−1 Whh f
+ bfh (20)
hbt = tanh xt Wxh
b
+ hbt−1 Whh

b
+ bbh (21)
ytb = ht Wo + bo (22)
hft hbt .
ht is composed of integrating and
IV. C ASE STUDY AND NUMERICAL RESULTS
In this section, the simulation environment applied in this
research and the details of datasets are explained firstly, and
a comprehensive prediction evaluation and comparison with
well-known benchmark models are performed subsequently.
A. Wind turbine simulation environment
Ct = gt  it + Ct−1  ft (18) In this work, the 5-MW reference horizontal axis wind
turbine of the National Renewable Energy Laboratory (NREL)
ht = tanh (Ct )  ot (19) is used as a full nonlinear aero-elastic model to perform
a broad range of wind speeds simulations [40]. Table II
where σ denotes the logistic sigmoid and tanh is the hy- shows the detailed characteristics of the proposed NREL wind
perbolic tangent activation function. W and b are the weight turbine.

1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
7

1) TurbSim wind field simulator: TurbSim is a full-field, uph×pv×pt , vph×pv×pt and wph×pv×pt in x, y, and z directions,
stochastic, turbulent-wind simulator developed by the NREL. respectively, where ph and pv denote the number of horizontal
It uses statistical models to generate realistic time series and vertical points in the spatial grid, respectively, and pt
of longitudinal, crosswise, and vertical components of the represents the temporal dimension of the vector. A broad
wind field [41]. Taylor’s frozen wind hypothesis is utilized range of the wind fields is generated according to various
to simulate wind signals; that is, wind field is modeled as a mean longitudinal wind speeds from 6 to 24 m/s with 2 m/s
turbulence box marching toward the turbine. Thus, the wind resolution steps to ensure that the forecasting model could be
fields do not evolve with time and are assumed to be frozen trained with different wind field characteristics. The complete
from the LIDAR focal point until the turbine. The height and data set contains 3000 wind fields and the corresponding
width of the wind field grid are chosen to be 145 m to be large electrical generator power and low-speed shaft torque. In this
enough to encompass the entire rotor disk of the proposed work, the training, validation and testing sets account for 75%,
wind turbine. Moreover, the hub is horizontally centered in the 15%, and 10% of the dataset, respectively [43].
grid, so the top of the grid can be determined by the turbine
hub height plus rotor radius. C. Evaluation Criteria for wind turbine response forecasting
2) FAST wind turbine simulator:: The fatigue, aerodynam-
ics, structures, and turbulence (FAST) code developed by In this study, the root means square error (RMSE), the mean
the NREL is used to simulate the response of the proposed absolute error (MAE), and the mean absolute percentage error
turbine by providing a full nonlinear and high-fidelity turbine (MAPE) are employed as three evaluation metrics to evaluate
response simulation [42]. The wind field vectors created by the forecasting performance of the different models as follows:
the TurbSim are applied to the FAST turbine simulator to
v
u
u1 X N
compute the turbine’s output signal and state vector. In this RM SE = t y (n) − y(n))
(b
2
(23)
study, electrical generator power (GenPwr) and low-speed N n=1
shaft torque (LSShftTq) are considered as the wind turbine
responses’ representatives. It should be noted that we enabled N
the degrees of freedom (DOFs) associated with first and 1 X
M AE = |b
y (n) − y(n)| (24)
second blade flap-wise modes (2 × 3 DOFs), first and second N n=1
tower side-to-side modes (2 DOFs), first and second tower
fore-aft modes (2 DOFs), first and second blade edgewise N  
1 X  yb (n) − y(n) 
mode (2 × 3 DOFs), drive-train mode (1 DOF), and generator M AP E =  × 100% (25)
mode (1 DOF). N n=1  yb (n)

where yb (n), y(n) and N represents the desired output, the

B. Data description actual output, and the number of samples, respectively.
The proposed approach aims to perform the ultra-short-
term forecasting of wind turbine responses by taking the 2-D- D. Results and comparisons
images of wind fields and hub-height wind speed time series
To verify the efficiency and validity of the proposed frame-
as inputs for the CNN and BLSTM networks, respectively.
work, several single and hybrid forecasting approaches that
This means that besides hub height wind speed time series,
have been proposed in the literature are chosen as the bench-
upcoming wind fields are also utilized to increase the capa-
marks. ARIMA, MLP, DBN, IPDL, GRU, LSTM, and BLSTM
bility of the proposed model on capturing complex wind data
models use 1D-wind speed time series, 2D-CNN use 2D-wind
abstractions.
fields images, and 2D-CNN-MLP, 2D-CNN-GRU, and 2D-
As mentioned earlier, the NREL TurbSim package is em-
CNN-LSTM use both of them similar to the proposed 2D-
ployed to generate turbulent wind fields and NREL FAST
CNN-BLSTM model. As these machine learning models are
turbine simulator is used to model high order aeroelastic
completely data-dependent, all hyperparameters, such as the
nonlinear wind turbine. As depicted in Fig. 6, each Turbsim
simulation generates 3×3-D wind field components vectors
TABLE III: Cross-Validation results of proposed model with
some typical BLSTM structures for 1-step ahead power fore-
TABLE II: Characteristics of NREL 5-MW Wind Turbine
casting.
Symbol Parameter Value
RMSE MAE MAPE Online execution time
BLSTM structure
Prated Rated power 5 MW (kW) (kW) (%) (ms)
Mrated Rated generator torque 43.1 kN m [100 100] 63.48 29.14 2.26 0.31
Vin Cut-in wind speed 3 m/s [250 250] 63.01 28.78 2.21 0.38
Vrated Rated wind speed 11.4 m/s [100 100 50] 58.13 25.64 1.93 0.46
Vout Cut-out wind speed 25 m/s [250 250 50] 56.27 23.77 1.92 0.61
hhub Hub height 90 m [250 250 100] 61.22 26.94 2.02 0.73
Rrotor Rotor radius 63 m [250 250 250] 61.38 27.18 2.11 0.81

1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
8

TABLE IV: Forecasting performance comparison for 1-step and 2-step ahead tasks.

1-Step (10 sec) 2-step (20 sec) 1-Step (10 sec) 2-step (20 sec)
Method
RMSE MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE MAE MAPE
(kW) (kW) (%) (kW) (kW) (%) (kN.m) (kN.m) (%) (kN.m) (kN.m) (%)
ARIMA [44] 351.80 219.17 21.76 463.82 424.34 32.43 591.51 326.42 32.27 746.58 717.64 43.85
DBN [45] 338.63 214.31 20.59 458.91 408.70 32.02 589.72 318.43 29.85 718.89 668.54 40.66
MLP [46] 325.68 196.64 19.32 424.78 391.36 27.56 560.59 295.98 27.49 650.23 577.73 34.54
GRU [47] 277.63 163.73 15.81 369.46 342.68 24.07 535.97 255.62 20.93 548.01 437.58 27.11
LSTM [15] 281.02 165.66 16.04 365.71 338.51 22.83 534.71 255.57 20.07 542.23 435.43 25.64
IPDL [48] 267.16 159.93 15.38 342.07 312.22 21.39 466.69 247.91 19.72 537.64 421.80 23.96
BLSTM [49] 263.01 154.38 15.29 330.08 298.21 21.01 454.46 239.34 17.96 518.56 398.44 23.80
2D-CNN 234.02 107.08 8.64 241.29 140.37 9.25 252.95 97.9 8.11 282.95 193.81 10.86
2D-CNN-MLP 176.23 77.31 6.61 190.57 139.49 8.49 197.17 87.94 7.66 234.86 185.55 9.94
2D-CNN-GRU 150.89 52.97 3.66 160.05 68.49 4.73 110.38 59.18 5.31 153.34 123.02 6.90
2D-CNN-LSTM 148.33 51.38 3.61 155.61 64.82 4.46 110.22 59.19 5.29 150.59 111.56 5.73
2D-CNN-BLSTM 58.34 24.73 1.98 67.02 41.05 2.74 102.61 51.22 4.55 145.11 105.58 5.41

Targets Targets
4000 4000
Forecasted Forecasted
Generated wind power (kW)

Generated wind power (kW)

3500 3500

3000 3000

2500 2500

2000 2000

1500 1500

0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time steps (10 sec) Time steps (10 sec)

(a) (b)

Targets Targets
4000 4000
Forecasted Forecasted
Generated wind power (kW)

Generated wind power (kW)

3500 3500

3000 3000

2500 2500

2000 2000

1500 1500

0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time steps (10 sec) Time steps (10 sec)

(c) (d)
Fig. 7: Wind power forecasting comparison of four 2D-CNN-based models for the first 60 steps (600 sec) of test data. (a)
2D-CNN. (b) 2D-CNN-MLP. (c) 2D-CNN-LSTM. (d) 2D-CNN-BLSTM.

number of neurons in each layer and the number of layers
in each model, are optimally tuned using the training data
according to common practice recommended by the deep
learning community [50]. The optimal number of hidden
layers for MLP model is determined to be 2 with 50 hidden
nodes in each layer. The DBN is established by using three
hidden layers with 100, 125, and 125 units, respectively. The
overall structure of the GRU, LSTM and B-LSTM networks
are composed of three main hidden layers with 250, 250,
and 50 units, respectively. The IPDL model consists of three
stacked layers with a linear regression model at the top.
The CNN structure comprises six layers, including three
convolution layers, three pooling layers, and a flattening layer.
To verify the performance of the validation, the 1-step ahead
power forecasting case study is repeated for different network
structures. Table III shows the cross-validation results of some
typical BLSTM structures with different number of layers
and blocks. As shown in this table, both the increase and
decrease in the network parameters reduces the performance
according to the forecasting indices. The online computation
time of the proposed method for different network parameters
is calculated in Table III. Considering that the sampling time
scale of the forecasting is 10 s or longer, the computational
time of around 0.6 ms is relatively fast enough to guarantee
reliable and safe wind turbine real-time operation. Moreover,
all models are implemented in Python with the Keras library
and TensorFlow as the backend [51]. The workstation used
is configured with an Intel Core TM i7-8700 3.2 GHz CPU,
NVIDIA GPU GeForce GTX 1070 GPU, and 32 GB of RAM.
Table IV compares the forecasting results and provides

1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
2000
Proposed
CNN-LSTM
1800
CNN-GRU
1600
CNN-MLP
CNN
1400
BLSTM
IPDL
LSTM
GRU
1200
RMSE (kW)
MLP
DBN
1000
ARIMA
800
600
400
200
0
1 2 3
Time Steps
4 5
Fig. 8: Comparison of power forecasting RMSE results for
1-step up to 6-step ahead tasks (step time = 10 sec).
the average test RMSE, MAE, and MAPE for ultra-short-
term generated power and load predictions with 1-step (10
sec) and 2-step (20 sec) ahead tasks, respectively. This table
clearly shows that the proposed method (2D-CNN-BLSTM)
outperforms the other benchmarking models and has the
best performance in both 1-step and 2-step ahead forecasting
tasks. The results show that machine learning models perform
better than ARIMA as a linear statistical model. DBN has
the largest error range and randomness, which is due to
the pretraining and supervised training process. The IPDL
model yields more accurate predictions when compared to
the classic DBN. According to the results, GRU and LSTM
approaches both outperform shallow architecture, MLP, since
these deep recurrent networks can better model the highly
nonlinear temporal features of wind time series. As the variant
of LSTM, the forecasting metrics of GRU are closer to those of
LSTM, but in 2-step ahead scenarios LSTM has a remarkable
improvement in all criteria. BLSTM is the best single time
series-based architecture compared to the LSTMS and MLP
models. Moreover, using wind fields has a significant effect on
the prediction accuracy which can be understood by comparing
first three models, MLP, LSTM, BLSTM, with other 2D-CNN-
based ones. For example, the RMSE of 2D-CNN-MLP for 1-
step and 2-step ahead wind power predictions are 176 kW and
190.57 kW, respectively, which are increased to 325.68 kW
and 424.78 kW for the single MLP model. MAPE result of
2D-CNN-BLSTM for 1-step predictions is 1.98 which reaches
to 2.74 in the 2-step compared to the best time series-based
approach, single BLSTM, which has MAPE of 15.29 and
21.01 in 1-step and 2-step, respectively. Therefore, applying
single time series-based models for longer-term predictions
cannot yield reliable performance. 2D-CNN-MLP has 24.69%
power RMSE and 22.05% load RMSE improvements over
single 2D-CNN. These improvements are further increased to
36.61% and 56.42% for the power RMSE and load RMSE
results, respectively, when the MLP is replaced by the deep
LSTM network. The more precise forecast shows the better
generalization capability of the deep recurrent models.
The proposed 2D-CNN-BLSTM obtains better results com-
pared to 2D-CNN-LSTM. 2D-CNN BLSTM outperforms 2D-
1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
9
CNN-LSTM with 51.86% and 13.46% MAE improvements
in power and load 1-step ahead forecasts. Furthermore, this
model decreases MAE by 36.67% and 5.36% for power and
load 2-step ahead prediction tasks, respectively. The higher
accuracy of 2D-CNN-BLSTM indicates the superiority of bidi-
rectional learning method to effectively capture the previous
and future hidden features of the proposed wind data.
To provide better visualization, Fig. 7 demonstrates the
wind power 1-step ahead prediction results of the four models
which use both 1D-wind speed time series and 2D-wind fields
as inputs. As shown in this figure, 2D-CNN is dominated
by 2D-CNN-MLP and 2D-CNN-LSTM has relatively better
6 performance in spike points than both of them. Our deep
hybrid model significantly outperforms all of them and can
follow the sharp spikes accurately.
Fig. 8 depicts the performance comparison of all single and
hybrid models for extended time horizons from 1-step ahead to
6-step ahead prediction tasks. For the extended prediction time
step, hybrid models with more complex structure are needed
to achieve high accuracy by using both 1-D and 2-D data.
As shown in Fig. 8, single architectures are dominated by
hybrid deep learning models in all time horizons. BLSTM and
IPDL have relatively good performance compared to CNN-
based models; however, for larger forecasting time steps, the
plot shows that hybrid frameworks outperform single ones.
Fig. 9 illustrates the regression responses of the proposed
method for wind output power and fatigue load as two
representatives of wind turbine responses. As can be seen in
this figure, the predicted values mostly stay below the blue line
and show slightly higher error in wind powers above 4500
kW and higher than 4000 kN.m fatigue loads. The fatigue
load case has faced some higher errors compared to the wind
power case which seems to be reasonable in that the fatigue
load is highly influenced by the dynamic interactions between
the turbine structure and the wind flow. Overall, as it is clear,
our proposed method, shows the acceptable performance in
both tasks and can helps us to make more accurate results.
To investigate the effect of wind fields snapshots on the
performance of our proposed model in another way, another
extension of 2D-CNN-BLSTM is designed as our baseline.
This baseline methodology only uses a longitudinal component
of the wind field, u, instead of entire wind field components.
Fig. 10 compares the 1-step ahead load prediction results of
proposed and baseline models. As it is clear, our proposed
method, by considering all wind fields components, shows
slightly better performance and generalization capability com-
pared to the baseline model which considers only streamwise
component, especially when load time series has an abrupt
change. It should be noted that the main limitation of this
work lies in the fact that these improvements are conditioned
to have an effective preview of wind data provided by the
LiDAR sensor, which is still a high-cost solution.
In case the LiDAR system is failed due to an unknown
reason and consequently 2D wind field images are not avail-
able, the proposed framework can still be used to forecast
as it is kind of ensemble learning and combined CNN and
BLSTM networks in a parallel manner. Fig. 11 compares the
wind power 1-step ahead prediction results of the three models
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
10

5000

4500 4000

4000
3500
Predicted values

Predicted values
3500

3000
3000

2500 2500

2000
2000

1500
1500
1000
1000 1500 2000 2500 3000 3500 4000 4500 5000 1500 2000 2500 3000 3500 4000 4500
Target Values Target Values

(a) (b)
Fig. 9: Regression plots of the proposed method (2D-CNN-BLSTM): (a) Wind power (kW); (b) fatigue load (kN.m).

4500 4500

4000 4000
Fatigue load (kN.m)

Fatigue load (kN.m)

3500 3500

3000 3000

2500 2500
Targets Targets
Forecasted Forecasted
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time steps (10 sec) Time steps (10 sec)

(a) (b)
Fig. 10: Fatigue load forecasting results comparison of proposed 2D-CNN-BLSTM and baseline model for the first 60 steps
(600 sec) of test data. (a) baseline model. (b) proposed 2D-CNN-BLSTM model.

(a) (b) (c)

Fig. 11: Wind power forecasting comparison of three Image-based models for the first 60 steps (600 sec) of test data in case
of LiDAR failure. (a) 2D-CNN-MLP. (b) 2D-CNN-LSTM. (c) 2D-CNN-BLSTM.

which previously used both 1D-wind speed time series and and HNN-QR models, MAPEs of power forecast are 4.42%
2D-wind fields as inputs. As shown in this figure, all of the and 4.07%, respectively. Comparing HNN-QR with 2D-CNN-
methods have relatively higher prediction error than the normal BLSTM, for the wind power prediction, 2D-CNN-BLSTM
situation (Fig. 7). The figure shows that the predicted values method can improve MAPE by 51.35%. Similar trend can be
of 2D-CNN-BLSTM generally conform well with the target seen for the load forecasting. 2D-CNN-BLSTM can forecast
values with low average error levels. the most accurate ultra-short-term power and load prediction
Fig. 12 compares the ultra-short-term generated power tasks, which can get the best forecasting metrics. This finding
and load 1-step forecasting performance of the proposed can verify the superiority of the proposed 2D-CNN-BLSTM
model with spatiotemporal prediction methods which have model.
been recently employed in the literature, including GDNN
[52], CNN-Taguchi [53], HNN-QR [31], and STNN-VB
V. C ONCLUSION
[54]. It is worth noting that the procedure of designing an
optimal structure for the other frameworks is similar to that This paper develops a novel LiDAR-assisted deep 2-D
of the proposed 2D-CNN-BLSTM model. In the generated CNN-BLSTM model for the ultra-short-term prediction of
power prediction task, CNN-Taguchi and STNN-VB produce future wind turbine responses using upcoming sequences of
the MAPE of 8.52% and 7.14%, respectively. For GDNN full wind field components and hub-height wind speed time

1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
10
Power Load
9 [5] A. Mahdizadeh, R. Schmid, and D. Oetomo, “LIDAR-Assisted Exact
8 Output Regulation for Load Mitigation in Wind Turbines,” IEEE Trans.
7
6
MAPE
5
4 [7] M. B. Ozkan and P. Karagoz, “A novel wind power forecast model:
3 Statistical hybrid wind power forecast technique (SHWIP),” IEEE Trans.
2
1
0
2D-CNN-BLSTM HNN-QR GDNN STNN-VB CNN-Taguchi
Forecasting method
Fig. 12: Comparison of power and fatigue load forecasting
results with different spatiotemporal models.
series before reaching the turbine blades as inputs. As a data-
driven framework, the performance of the proposed model
is determined solely by the potential interactions hidden in
the wind field and time series data rather than the physical
equations or predetermined distribution types. Thus, it can
avoid the dual risks of model incorrectness or distribution type
misspecification. The NREL 5-MW reference horizontal axis
wind turbine with FAST are utilized for simulations. Realistic
3-D wind field components vectors are generated by NREL
TurbSim. The proposed 2-D CNN-BLSTM model is designed
for ultra-short-term forecasting of wind turbine response and
shows the high-quality outputs with the smallest metrics. For
example, it has demonstrated 78% and 75% improvement in
RMSE when compared with single BLSTM and 2-D CNN
models, respectively.
The proposed model employs 2D-CNN and BLSTM net-
works to better handle complex spatial- temporal features
from the highly variable wind data compared to conventional
forecasting methods which simply use historical time series
data. The main advantage of proposed model over other
deep learning-based forecasting methods is that it uses wind
preview information provided by LIDAR as an advanced
remote sensing wind measurement technology. Thus, it can
be helpful for the wind farm operators as an efficient tool
in yaw misalignment and de-loading control strategies. While
this study is geared towards exploring the effects of the
proposed forecasting methodology on onshore wind turbine
control strategies, future work could study the effectiveness of
adding LIDAR to offshore wind turbines and considering the
sea current and wave data as external inputs.
R EFERENCES
[1] “Renewable Energy in Alberta.” [Online]. Available: https://www.
alberta.ca/renewable-energy-in-alberta.aspx/.
[2] A. Dolatabadi, B. Mohammadi-Ivatloo, M. Abapour, and S. Tohidi,
“Optimal stochastic design of wind integrated energy hub,” IEEE Trans.
Ind. Inf., vol. 13, no. 5, pp. 2379–2388, 2017.
[3] A. Dolatabadi, M. Jadidbonab, and B. Mohammadi-ivatloo, “Short-term
scheduling strategy for wind-based energy hub: A hybrid stochastic/igdt
approach,” IEEE Trans. Sustainable Energy, vol. 10, no. 1, pp. 438–448,
2019.
1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
11
[4] Q. Yao, J. Liu, and Y. Hu, “Optimized active power dispatching strategy
considering fatigue load of wind turbines during de-loading operation,”
IEEE Access, vol. 7, pp. 17 439–17 449, 2019.
Control Syst. Technol., 2020.
[6] M. Khodayar, O. Kaynak, and M. E. Khodayar, “Rough deep neural
architecture for short-term wind speed forecasting,” IEEE Trans. Ind.
Inf., vol. 13, no. 6, pp. 2770–2779, 2017.
Ind. Inf., vol. 11, no. 2, pp. 375–387, 2015.
[8] S. Buhan and I. Çadırcı, “Multistage wind-electric power forecast by
using a combination of advanced statistical methods,” IEEE Trans. Ind.
Inf., vol. 11, no. 5, pp. 1231–1242, 2015.
[9] H. B. Azad, S. Mekhilef, and V. G. Ganapathy, “Long-term wind speed
forecasting and general pattern recognition using neural networks,” IEEE
Trans. Sustainable Energy, vol. 5, no. 2, pp. 546–553, 2014.
[10] D. Allen, A. Tomlin, C. Bale, A. Skea, S. Vosper, and M. Gallani,
“A boundary layer scaling technique for estimating near-surface wind
energy using numerical weather prediction and wind map data,” Appl.
Energy, vol. 208, pp. 1246–1257, 2017.
[11] P. Chen, T. Pedersen, B. Bak-Jensen, and Z. Chen, “ARIMA-based time
series model of stochastic wind power generation,” IEEE Trans. Power
Syst., vol. 25, no. 2, pp. 667–676, 2009.
[12] K. Bhaskar and S. Singh, “AWNN-assisted wind power forecasting using
feed-forward neural network,” IEEE Trans. Sustainable Energy, vol. 3,
no. 2, pp. 306–315, 2012.
[13] Z. Shi, H. Liang, and V. Dinavahi, “Direct interval forecast of uncertain
wind power based on recurrent neural networks,” IEEE Trans. Sustain-
able Energy, vol. 9, no. 3, pp. 1177–1187, 2017.
[14] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
Comput., vol. 9, no. 8, pp. 1735–1780, 1997.
[15] Y.-X. Wu, Q.-B. Wu, and J.-Q. Zhu, “Data-driven wind speed forecasting
using deep feature extraction and LSTM,” IET Renewable Power Gener.,
vol. 13, no. 12, pp. 2062–2069, 2019.
[16] M.-R. Chen, G.-Q. Zeng, K.-D. Lu, and J. Weng, “A two-layer nonlinear
combination method for short-term wind speed prediction based on
ELM, ENN, and LSTM,” IEEE Internet Things J., vol. 6, no. 4, pp.
6997–7010, 2019.
[17] M. Harris, M. Hand, and A. Wright, “Lidar for turbine control,” National
Renewable Energy Laboratory, Golden, CO, Report No. NREL/TP-500-
39154, 2006.
[18] L. Zhang and Q. Yang, “A Method for Yaw Error Alignment of Wind
Turbine Based on LiDAR,” IEEE Access, vol. 8, pp. 25 052–25 059,
2020.
[19] D. Schlipf, P. Fleming, F. Haizmann, A. Scholbrock, M. Hofsäß,
A. Wright, and P. W. Cheng, “Field testing of feedforward collective
pitch control on the CART2 using a nacelle-based lidar scanner,” in J.
Phys. Conf. Ser., vol. 555, no. 1. IOP Publishing, 2014, p. 012090.
[20] A. Clifton, P. Clive, J. Gottschall, D. Schlipf, E. Simley, L. Simmons,
D. Stein, D. Trabucchi, N. Vasiljevic, and I. Würth, “IEA Wind Task 32:
Wind lidar identifying and mitigating barriers to the adoption of wind
lidar,” Remote Sensing, vol. 10, no. 3, p. 406, 2018.
[21] M. A. Mohandes and S. Rehman, “Wind speed extrapolation using
machine learning methods and LiDAR measurements,” IEEE Access,
vol. 6, pp. 77 634–77 642, 2018.
[22] D. Qin, J. Yu, G. Zou, R. Yong, Q. Zhao, and B. Zhang, “A novel
combined prediction scheme based on cnn and lstm for urban pm 2.5
concentration,” IEEE Access, vol. 7, pp. 20 050–20 059, 2019.
[23] J. Li, X. Li, and D. He, “A directed acyclic graph network combined
with cnn and lstm for remaining useful life prediction,” IEEE Access,
vol. 7, pp. 75 464–75 475, 2019.
[24] Q. Wang, S. Bu, Z. He, and Z. Y. Dong, “Toward the prediction level of
situation awareness for electric power systems using cnn-lstm network,”
IEEE Trans. Ind. Inf., 2020.
[25] L. Ren, J. Dong, X. Wang, Z. Meng, L. Zhao, and J. Deen, “A data-
driven auto-cnn-lstm prediction model for lithium-ion battery remaining
useful life,” IEEE Trans. Ind. Inf., 2020.
[26] X. Song, F. Yang, D. Wang, and K.-L. Tsui, “Combined cnn-lstm
network for state-of-charge estimation of lithium-ion batteries,” IEEE
Access, vol. 7, pp. 88 894–88 902, 2019.
[27] T.-Y. Kim and S.-B. Cho, “Predicting residential energy consumption
using cnn-lstm neural networks,” Energy, vol. 182, pp. 72–81, 2019.
[28] F. U. M. Ullah, A. Ullah, I. U. Haq, S. Rho, and S. W. Baik, “Short-term
prediction of residential power energy consumption via cnn and multi-
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2021.3097716, IEEE
Transactions on Industrial Informatics
layer bi-directional lstm networks,” IEEE Access, vol. 8, pp. 123 369–
123 380, 2019.
[29] K. Wang, X. Qi, and H. Liu, “Photovoltaic power forecasting based
lstm-convolutional network,” Energy, vol. 189, p. 116225, 2019.
[30] H. Zang, L. Liu, L. Sun, L. Cheng, Z. Wei, and G. Sun, “Short-term
global horizontal irradiance forecasting based on a hybrid cnn-lstm
model with spatiotemporal correlations,” Renewable Energy, vol. 160,
pp. 26–41, 2020.
[31] Y. Yu, X. Han, M. Yang, and J. Yang, “Probabilistic prediction of
regional wind power based on spatiotemporal quantile regression,” IEEE
Trans. Ind. Appl., vol. 56, no. 6, pp. 6117–6127, 2020.
[32] M. Afrasiabi, M. Mohammadi, M. Rastegar, and S. Afrasiabi, “Advanced
deep learning approach for probabilistic wind speed forecasting,” IEEE
Trans. Ind. Inf., vol. 17, no. 1, pp. 720–727, 2020.
[33] Z. Wang, J. Zhang, Y. Zhang, C. Huang, and L. Wang, “Short-term wind
speed forecasting based on information of neighboring wind farms,”
IEEE Access, vol. 8, pp. 16 760–16 770, 2020.
[34] Y. Ju, G. Sun, Q. Chen, M. Zhang, H. Zhu, and M. U. Rehman, “A
model combining convolutional neural network and lightgbm algorithm
for ultra-short-term wind power forecasting,” IEEE Access, vol. 7, pp.
28 309–28 318, 2019.
[35] Q. Zhu, J. Chen, D. Shi, L. Zhu, X. Bai, X. Duan, and Y. Liu, “Learning
temporal and spatial correlations jointly: A unified framework for wind
speed prediction,” IEEE Trans. Sustainable Energy, vol. 11, no. 1, pp.
509–523, 2019.
[36] A. Peña, C. B. Hasager, J. Lange, J. Anger, M. Badger, F. Bingöl,
O. Bischoff, J.-P. Cariou, F. Dunne, S. Emeis et al., Remote sensing
for wind energy. DTU Wind Energy, 2013.
[37] P. A. Fleming, A. Scholbrock, A. Jehu, S. Davoust, E. Osler, A. D.
Wright, and A. Clifton, “Field-test results using a nacelle-mounted
lidar for improving wind turbine power capture by reducing yaw
misalignment,” in J. Phys. Conf. Ser., vol. 524, no. 1. IOP Publishing,
2014, p. 012002.
[38] A. D. Hansen, F. Iov, P. E. Sørensen, N. A. Cutululis, C. Jauch, and
F. Blaabjerg, “Dynamic wind turbine models in power system simulation
tool DIgSILENT,” 2007.
[39] H. Samet, S. Ketabipoor, M. Afrasiabi, S. Afrasiabi, and M. Moham-
madi, “Deep Learning Forecaster based Controller for SVC: Wind Farm
Flicker Mitigation,” IEEE Trans. Ind. Inf., 2020.
[40] J. Jonkman, S. Butterfield, W. Musial, and G. Scott, “Definition of a 5-
MW reference wind turbine for offshore system development,” National
Renewable Energy Lab.(NREL), Golden, CO (United States), Tech.
Rep., 2009.
[41] B. J. Jonkman, “TurbSim user’s guide: Version 1.50,” National Renew-
able Energy Lab.(NREL), Golden, CO (United States), Tech. Rep., 2009.
[42] J. M. Jonkman and M. L. Buhl Jr, “FAST user’s guide-updated august
2005,” National Renewable Energy Lab.(NREL), Golden, CO (United
States), Tech. Rep., 2005.
[43] H. Jahangir, H. Tayarani, S. Sadeghi Gougheri, M. Aliakbar Golkar,
A. Ahmadian, and A. Elkamel, “Deep Learning-based Forecasting Ap-
proach in Smart Grids with Micro-Clustering and Bi-directional LSTM
Network,” IEEE Trans. Ind. Electron., pp. 1–1, 2020.
[44] P. Gangwar, A. Mallick, S. Chakrabarti, and S. N. Singh, “Short-term
forecasting-based network reconfiguration for unbalanced distribution
systems with distributed generators,” IEEE Trans. Ind. Inf., vol. 16, no. 7,
pp. 4378–4389, 2019.
[45] K.-P. Lin, P.-F. Pai, and Y.-J. Ting, “Deep belief networks with genetic
algorithms in forecasting wind speed,” IEEE Access, vol. 7, pp. 99 244–
99 253, 2019.
[46] L. Wang, Z. Zhang, H. Long, J. Xu, and R. Liu, “Wind Turbine Gearbox
Failure Identification With Deep Neural Networks,” IEEE Trans. Ind.
Inf., vol. 13, no. 3, pp. 1360–1368, 2017.
[47] C. Li, G. Tang, X. Xue, A. Saeed, and X. Hu, “Short-term wind
speed interval prediction based on ensemble GRU model,” IEEE Trans.
Sustainable Energy, vol. 11, no. 3, pp. 1370–1380, 2020.
[48] M. Khodayar, J. Wang, and M. Manthouri, “Interval deep generative
neural network for wind speed forecasting,” IEEE Trans. Smart Grid,
vol. 10, no. 4, pp. 3974–3989, 2019.
[49] A. Dolatabadi, H. Abdeltawab, and Y. A.-R. I. Mohamed, “Hybrid deep
learning-based model for wind speed forecasting based on dwpt and
[51] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin,
S. Ghemawat, G. Irving, M. Isard et al., “Tensorflow: A system for large-
1551-3203 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on July 29,2021 at 18:40:20 UTC from IEEE Xplore. Restrictions apply.
12
bidirectional lstm network,” IEEE Access, vol. 8, pp. 229 219–229 232,
2020.
[50] I. Goodfellow, Y. Bengio, A. Courville, and Y. Bengio, Deep learning.
MIT press Cambridge, 2016, vol. 1.
scale machine learning,” in 12th {USENIX} symposium on operating
systems design and implementation ({OSDI} 16), 2016, pp. 265–283.
[52] M. Khodayar and J. Wang, “Spatio-temporal graph deep neural network
for short-term wind speed forecasting,” IEEE Trans. Sustainable Energy,
vol. 10, no. 2, pp. 670–681, 2018.
[53] Y.-Y. Hong and T. R. A. Satriani, “Day-ahead spatiotemporal wind speed
forecasting using robust design-based deep learning neural network,”
Energy, vol. 209, p. 118441, 2020.
[54] Y. Liu, H. Qin, Z. Zhang, S. Pei, Z. Jiang, Z. Feng, and J. Zhou, “Prob-
abilistic spatiotemporal wind speed forecasting based on a variational
bayesian deep learning model,” Applied Energy, vol. 260, p. 114259,
2020.
Amirhossein Dolatabadi (GS’16) was born in
Tabriz, Iran in 1991. He received the B.Sc. and
M.Sc. degrees in electrical engineering from Uni-
versity of Tabriz, Tabriz, Iran, in 2014 and 2016,
respectively. From 2016 to 2018, he served as a
Research Assistant with the Smart Energy Systems
Laboratory (SES Lab), University of Tabriz.
He is currently pursuing the PhD degree at the
University of Alberta, Edmonton, Canada. His re-
search interests include artificial intelligence, ma-
chine learning and data analytics applications in
energy systems.
Hussein Abdeltawab (GS’12–M’17) was born in
Bani-Souwaif, Egypt, in April 1987. He received
the B.Sc. (Hons.) and M.Sc. degrees in electrical
engineering from Cairo University, in 2009 and
2012, respectively, and the Ph.D. degree in electrical
engineering from the University of Alberta, Edmon-
ton, AB, Canada, in 2017.
He is currently an Assistant Professor of electrical
engineering with Penn State University, Erie, PA,
USA. He is also a licensed Professional Engineer
in Saskatchewan, Canada. His research interests in-
clude energy management, control system applications in renewable energy,
energy storage, and smart distribution systems.
Yasser Abdel-Rady I. Mohamed (M’06–SM’11–
F’021) was born in Cairo, Egypt, on November 25,
1977. He received the B.Sc. (with honors) and M.Sc.
degrees in electrical engineering from Ain Shams
University, Cairo, in 2000 and 2004, respectively,
and the Ph.D. degree in electrical engineering from
the University of Waterloo, Waterloo, ON, Canada,
in 2008.
He is currently with the Department of Electrical
and Computer Engineering, University of Alberta,
AB, Canada, as a Professor. His research interests
include dynamics and controls of power converters; grid integration of dis-
tributed generation and renewable resources, microgrids, modeling, analysis,
and control of smart grids; and electric machines and motor drives.
Dr. Mohamed is an Associate Editor of the IEEE Transactions on Power
Electronics and an Editor of the IEEE Transactions on Power Systems, IEEE
Transactions on Smart Grid, and IEEE Power Engineering Letters. He is a
registered Professional Engineer in the Province of Alberta.