See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/340166656

Predicting Wind Turbine Blade Erosion using Machine Learning

Article · September 2019

CITATIONS READS
12 1,752

5 authors, including:

James Herford
Walmart Stores
5 PUBLICATIONS 12 CITATIONS

SEE PROFILE

All content following this page was uploaded by James Herford on 25 March 2020.

The user has requested enhancement of the downloaded file.

SMU Data Science Review
Volume 2 | Number 2 Article 17

2019

Predicting Wind Turbine Blade Erosion using

Machine Learning
Casey Martinez
Southern Methodist University, [email protected]

Festus Asare Yeboah

Southern Methodist University, [email protected]

Scott Herford
Southern Methodist University, [email protected]

Matt Brzezinski
[email protected]

Viswanath Puttagunta
[email protected]

Follow this and additional works at: https://scholar.smu.edu/datasciencereview

Part of the Applied Statistics Commons, Engineering Commons, and the Longitudinal Data
Analysis and Time Series Commons

Recommended Citation
Martinez, Casey; Asare Yeboah, Festus; Herford, Scott; Brzezinski, Matt; and Puttagunta, Viswanath (2019) "Predicting Wind Turbine
Blade Erosion using Machine Learning," SMU Data Science Review: Vol. 2 : No. 2 , Article 17.
Available at: https://scholar.smu.edu/datasciencereview/vol2/iss2/17

This Article is brought to you for free and open access by SMU Scholar. It has been accepted for inclusion in SMU Data Science Review by an
authorized administrator of SMU Scholar. For more information, please visit http://digitalrepository.smu.edu.
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

Predicting Wind Turbine Blade Erosion using

Machine Learning

Casey R. Martinez, Festus A. Yeboah, James S. Herford,

Matt Brzezinski, Viswanath Puttagunta
1
Master of Science in Data Science
Southern Methodist University
Dallas, Texas USA
{casianom,fasareyeboah,jherford}@smu.edu,
[email protected], [email protected]

Abstract. Using time-series data and turbine blade inspection assess-

ments, we present a regression model used to predict remaining turbine
blade life in wind turbines. Capturing the kinetic energy of wind requires
complex mechanical systems, which require sophisticated maintenance
and planning strategies. There are many traditional approaches to mon-
itoring the internal gearbox and generator, but the condition of turbine
blades can be difficult to measure and access. Accurate and cost-effective
estimates of turbine blade life cycles drive optimal investments in repairs
and improve overall performance. These measures drive down costs, as
well as, provide cheap and clean electricity for the planet. It has been
shown that by sampling actual blade conditions with drone photography,
the remaining turbines can be predicted with acceptable accuracy(86%
R-Squared Error). This result not only saves money on drone inspections
but also allows for a more precise estimation of lost revenue and a clearer
understanding of repair investments.

Keywords: Wind Energy, Predictive Maintenance, Blade Erosion, Machine

Learning

1 Introduction

As a form of solar energy, wind is caused by three concurrent events: the un-
even heating of Earth’s atmosphere by the sun, irregularities found on Earth’s
surface, and the rotation of the Earth. Across the United States, the variation
of wind flow patterns and speeds are substantial and greatly modified by bodies
of water, vegetation, and differences in terrain. Both ”wind energy” and ”wind
power” describe the overall process where wind is leveraged for the purpose of
generating mechanical power or electricity; the kinetic energy harnessed from
wind turbines is thereby transformed into mechanical resulting in specific tasks,
such as grinding grain, pumping water, or using a generator to convert mechan-
ical power into electricity [1].

Published by SMU Scholar, 2019 1

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

The production of electricity from energy captured by wind turbines stems

from the aerodynamic force created by rotor blades. As wind flows across the
wind turbine blades, air pressure on one side of the blade decreases. The differ-
ence in air pressure across the two sides of the blade creates both lift and drag
[1]. Because the force of the lift is stronger than the drag, the rotor shaft is then
forced to spin. The generator contained within the turbine is connected to the
rotor shaft.

Fig. 1. Example of a Wind Turbine (Energy.gov)

Modern-day wind turbines are complex and sophisticated machines with hun-
dreds of moving parts. A team of technicians and engineers are needed to main-
tain operation at levels needed to provide cheap electricity to market. Most of
the rotating components inside the nacelle portion of the turbine have tempera-
ture and vibration sensors. These sensors alert the control system of anomalous
equipment behavior. Since the fiberglass blades themselves have no moving parts
and can be 50 meters in length or greater, similar monitoring systems have not
been implemented at scale. The industry relies on periodic inspection via tele-
scoping photography, drone photography, and/or human inspection with ropes
and harnesses. Each of these approaches require turbine downtime, which can
require labor intensive tasks, such as collecting large volumes of data and mak-
ing blade assessments. An example of traditional blade inspection performed by
hand with a technician rappelling down each blade (Figure 2).

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 2
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

Fig. 2. Example of a Manual Blade Inspection

Since it is not cost effective to inspect every turbine in the fleet, only a sample
of turbines are inspected. These inspections can also be prone to human error.
This error is due to the subjective nature of the assessments, which are usually
based on photography. The defect of interest in this work is the erosion of blade
materials on the leading edge. This failure mode is a slow process which can
etch away the gel coating finish of the fiberglass blade. Other defects caused
by manufacturing defects, lightning strikes, and other events are not considered
here.
There are numerous defect types that can appear on a blade from cracks and
pitting to full delamination in sections. The focus of this paper is on leading
edge erosion. Most of these defects are either caused by an acute event such
as lightning, blade erosion is caused by small physical impacts over time. This
defect type can be difficult to see in early stages and can have a significant impact
on power generation. Sources of particles in the air can be dust, rain, bugs, and
hail. Leading edge blade erosion caused by water droplet impingement reduces
blade aerodynamics and limits power output. If left uncorrected, the structural
integrity of the blade can be compromised.

2 Background
Leading-edge blade erosion has become a common problem among wind farm
operators. Better prediction of blade erosion, so maintenance can be planned and
scheduled, can produce a positive financial impact[2]. The entire wind energy sec-
tor is faced with effects from the erosion of materials used in wind turbine blades,

Published by SMU Scholar, 2019 3

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

which are primarily due to the effect of weather conditions[3]. Blade material
erodes due to the effect of rain, hailstones, and sand dust, which significantly
reduces the service life of wind turbines. Accelerated replacement of turbines
becomes expensive: up to 2 to 4 percent of the value of all wind-generated power
is lost as a result of this problem[3].
The blades consist of a fiberglass frame and shell with a gel coating on the
surface. These materials are used for their desirable weight to strength ratio. This
gel coating can be damaged by debris in the air such as dust, rain, hail, and other
materials. This is due to the high tip speed of the blades. For an average wind
speed of 13 m/sec, the tip of the blade can move at speeds approximately 120
mph. At this speed, the gel coating can be eroded over time. Figure 3 shows an
example of leading edge erosion.

Fig. 3. Example of Leading Edge Erosion

The blades of the turbine are the most difficult part of the turbine to collect
data from. Modern turbines have sophisticated condition monitoring systems
for the rotating equipment inside the nacelle. The expensive rotating equip-
ment, such as the generator and gear box, are monitored by temperature and
vibration sensors where damage can be flagged quickly. No direct measurements
of the blade aerodynamics or structural integrity is collected by the turbine.
Turbine damage has to reach a significant level before a loss in aerodynamics,
and therefore power efficiency, can be detected. Analytical methods to use this
reduction in turbine efficiency will be discussed in a later section.

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 4
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

Blade inspections have seen a rapid advancement due to autonomous drones

and burst photography. In some estimates, 85 percent of blades currently in-
stalled have defects[4]. Wind farm operators need a way to prioritize and esti-
mate repair costs. As can be seen in figure 4 below, automated blade inspection
based on drone photography can reduce human error and lead to more consistent
results.

Fig. 4. Example of a Blade Inspection Report

The blade inspection ratings, physical damage description, and correspond-

ing impacts to turbine performance can be see in Table 1. Turbine performance
is not affected until a score of three and above. If this change in turbine aerody-
namic efficiency can be directly observed in the turbine’s performance, perhaps
it can be used to help predict the turbine damage. This under-performance can
be estimated using the first year of a turbine’s operational life as a baseline.
This baseline assumes pristine condition of the blades and that the turbine has
seen a wide range of environmental conditions. Then, if erosion does exist, the
gradual drop in performance will be revealed by the divergence of the actual
and predicted performance, called residual. This performance residual is a key
feature to predicting blade erosion.
Traditionally, turbine performance was analyzed by careful study of the tur-
bine’s power curve. This is the relationship between wind speed and electrical
power output. The theoretical power curve is modeled by the turbine supplier
using the turbine power generation equation. This simple representation does
take air density into account but does not include other factors such as air speed
uniformity, known as turbulence, and vertical air flow, known as sheer. The air

Published by SMU Scholar, 2019 5

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

density, turbulence, and sheer can be calculated using the meteorological tower
data installed at the wind farm[5].
Factors such as terrain, wind direction, wind sheer, turbulence, air density
and waking effects of other turbines can cause a significant amount of noise in
the power curve. Accurate modeling of turbine power performance cannot be
properly captured in this two-dimensional power curve of wind speed and power
generation. A multivariate regression modeling approach is used.

3 Literature Review

Over the last few decades, various methods have been employed to predict
the performance and condition of wind turbine monitoring systems; however, the
majority of faults continue to be detected through planned maintenance sessions
[6]. Several studies have used Artificial Neural Networks (ANN) to model the
normal behavior of wind turbines; thus, ANNs are being used to monitor per-
formance in real time and predict wind power performance, which can leveraged
for the purpose of optimizing fault detection systems [6]. Costa et al. provide
several models that have been used extensively for short-term prediction [7].
For example, they present use cases of multiple time series and neural network
methods in order to output energy demand as well as energy product estimates.
According to Costa et al., these methods can be purposed for bringing significant
operational advantages to energy production grids like wind turbine farms [7].
As mentioned above, the SCADA system has been utilized for collecting and
storing data from wind farm operations. In order to manage power consumption
and turbine efficiency, the SCADA system was integrated with turbine control
systems to detect downtimes [8]. Further, data from SCADA has been used for
classifying turbine failure events into categories separated by level of severity
using statistical methods to determine reliability of wind farms [9]. D. Fawzy et
al. leveraged the Wind Turbine Erosion Predictor (WTEP) to predict erosion
rates of wind turbines while accounting for environmental factors [10].

4 Data Formatting

The data collected for this paper consists of three types: blade inspection
data, turbine performance data, and meteorological data.

4.1 Blade Inspection Data

Blade inspection data was collected via drone photography by two contract-
ing companies. Each blade was imaged on all sides after the blade was stopped
and moved to the vertical position. These images were later processed and au-
tomatically classified using an image classification AI technique. We believe the
classification uses contrasting pixels. All scores above three are reviewed by a
human to check for proper scoring and to provide additional commentary.

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 6
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

The altitude of the drone was collected during the inspection. Since the blade
is locked in the vertical position during inspection, the altitude corresponds to
the position along the length of the blade. The location of the defect along the
blade length is important to target different types of blade defects.
Table 1 shows how defects are scored and if turbine performance is impacted
due to changes in the aerodynamic efficiency of the blade. As you can see, when
the gel coat and fiberglass of the blade is damaged, the performance of the blade
is reduced. Level five damage can cause blades to crack and break if not repaired.
Such broken blades can strike the tower and cause the entire turbine to fall to
the ground.

Table 1. Overview of Inspection Scores

Score Description Performance Impact?

1 Cosmetic Damage No
2 Some Spotting No
3 Gel Coat Damage Yes
4 Fiberglass Damage Yes
5 Structural Damage Yes

According to the pie graph in figure 5, 85% of the defect data consists of
level one and two defects.

Fig. 5. Distribution of Severity Classes

Published by SMU Scholar, 2019 7

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

The inspection data used in this paper consists of approximately 84,000 de-
fects collected from 1,084 turbines over 17 wind farms across North America (see
Fig.6).

Fig. 6. Map of Wind Farms

Since the wind farm operators cannot afford to have every turbine inspected,
data was collected for only a sample of the turbines in the fleet. Also, over time,
different companies were hired for inspections making combining results difficult.
By consulting with a turbine engineer, a common defect score was developed.

Identifying the specific defect signature of blade erosion is very important

for accurate predictions. The inspection data contained records that did not
match the erosion failure we were interested in. Cosmetic defects and damage
caused by lightning strikes are considered noise as these do not cause performance
degradation in the turbine. In addition to the location of defects along the length
of the blade, the location along with width is also analyzed. Defects on the
leading edge of the blade and located from the tip of the blade to half the length
of the turbine blade. Figure 7 illustrates this by showing the distribution of
defects that occur along the blade. This defect filtering is critical to accurately
predicting leading-edge erosion.

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 8
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

Fig. 7. Defect Location along Blade Length

Fig. 8. Turbine Map by Severity

4.2 Turbine Performance Data

Modern wind turbines have many moving parts in the nacelle and almost
all of them have either a temperature or vibration sensor installed. Turbine
controllers use this data to optimize performance, decrease output in high stress
situations, or shut down for protection. The high resolution data is collected
and stored by SCADA systems as 10 minute average or one minute snap shot
records. The turbine data that is most useful in this analysis is power output and
wind speed. These two measurements are key to determining if the turbine is not
performing at expected efficiency. Other turbine data is used for fatigue analysis
where time spent in certain high stress conditions can be aggregated. This allows
for turbines to be compared in a snapshot in time with blade inspections.

Published by SMU Scholar, 2019 9

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

4.3 Meteorological data

Weather data can be useful to calculate wind turbulence, wind sheer, precip-
itation, and temperature. Most wind farms have at least one met tower where
these instruments are installed. For precipitation data, local airports or other
National Weather Service data sources may be needed.

Fig. 9. Example of Meteorological Tower

5 Solution Approach

At first glance the inspection data suggests using a time series model for this
problem; however, time series data is a sequence of data points that measures
the same metric over a consistent period of time. Data that is collected based on
the occurrence of an event is not considered to be time series data. The turbine
inspection data has sparse and inconsistent records on inspection and is based
on when the turbine inspection was conducted.
To use a time series approach for our data, means employing one of the
many interpolation formulas to convert the data into equally spaced observa-
tions. However, there are challenges to using this approach. The inspection data
comes from several sites with different time resolution (some monthly and others
quarterly). The usual approach to interpolate this type of data, is to sample at
the smallest time resolution, which will end up with a time series that is sampled

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 10
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

each month, but is actually recorded every quarter. This can affect our accuracy
negatively.
Instead of a time series approach, with data that has the granularity of the
time stamp for when the inspection was done, a snapshot approach was em-
ployed. The data was grouped at the turbine level and for each turbine features
engineered to contain aggregate of fatigue metrics over the turbine’s entire life.
When it comes to modeling data, we could not rely on the results of an
individual model because of the commonality that individual models suffer far
too much from bias and/or variances; thus, we decided to incorporate ensemble
learning, which makes predictions based on several different models. Combining
each model results in the ensemble model being more flexible which creates less
bias and less variance. The model becomes less data-sensitive[11].
Two of the most popular methods ensemble learning are bagging and boost-
ing. Bagging encompasses the training of several individual models in parallel
where each model is trained by a random subset of data. Boosting requires the
training of several individual models in sequence where each model learns from
the errors made by the previous model[11]. In the case of prediction-based prob-
lems involving small-to-medium structured/tabular data, the preferred approach
is utilizing decision-tree based algorithms[12]. XGBoost is a decision-tree-based
algorithm which falls under the ensemble learning method that uses a gradient
boosting framework[12].

6 Analysis

6.1 Feature Engineering

The most important feature to create is the blade tip speed. Since the data set
consists of turbines of various blade lengths and rotational speeds, this feature
will allow all turbines to have a comparable metric for modeling. Blade tip speed
is calculated according to Equation 1

v = ω ∗ r, (1)
where v is the linear speed at the blade tip (m/s), ω is the rotational speed
(radians/minute), and r is the blade length (meters).

Figure 10 shows an example of tip speed plotted by time for one turbine. You
can see the plateau of speed is reached when wind speed is above a threshold.
This rated speed varies by turbine model. Many other features are also used
including wind turbulence, wind sheer, and the product of blade tip speed and
rainfall.
These features are aggregated over time between the date of inspection and
the date of commissioning. These final aggregated features are the inputs to the
final regression model to predict the turbine’s blade condition. An example of
some of these features over time can be seen in figure 11.

Published by SMU Scholar, 2019 11

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

Fig. 10. Tip Speed over Time

Fig. 11. Fatigue Metrics Example

6.2 Machine Learning Model

The goal of the machine learning model is to predict the severity score of
turbines using the aggregated features created to capture blade fatigue. This
model is a boosted random forest regressor with ten-fold cross validation. Figure
12 shows the individual turbine results colored by site.

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 12
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

Fig. 12. Actual vs. Predicted by Site

The model residual analysis shows a well-defined normal distribution around

zero as seen in Fig. 13. This suggests low bias overall but further analysis is
needed to verify.

Published by SMU Scholar, 2019 13

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

Fig. 13. Model Residual Distribution

The residual bias in the RainWindLog feature was investigated in Fig. 14.
There is no clear pattern to the residuals which is a good sign of low bias.

Fig. 14. Model Residual by Feature

There does exist bias in the residual plot of severity shown in Fig. 15. There
is a linear relationship for severity below 2.8 where very low severity shows an
overestimated prediction and higher severity shows an underestimated predic-
tion. At very high severity, there is a bias of underestimated predictions for most
points above 3.3. The cause of this residual distribution is not clear.

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 14
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

Fig. 15. Model Residual by Severity

The residual analysis by wind farm also shows non-uniform residuals where
there is strong bias in both directions depending on the wind farm site. There
are numerous potential reasons for this behaviour. First, each site has unique
weather complexities that can skew the fatigue metrics. Second, the turbine
performance data set used in this analysis is not complete for all wind farms.
This could skew the aggregations and can be address with proper imputation.
Third, the amount of blade damage varies by site and could cause, or be caused,
by the residual by severity analysis discuss earlier. For example, K2 Wind has
the largest severity values in the data set (Fig.12). This large underestimation
of predicted scores matches both analysis.

Published by SMU Scholar, 2019 15

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

Fig. 16. Model Residual by Site

6.3 Feature Importance

A benefit to modelling with gradient boosting is the fact that it is able

to automatically provide estimates of features that contributed greatly to the
model. A feature is considered important if modifying its values significantly
changes the model’s predictability. As shown in fig 12, Rainwindlog came out
as a significant feature in blade erosion prediction. This feature accounts for the
tip speed of the blade, the length of the turbine blade length which varies across
location and the amount of rain. Rainwindlog coming out as the best feature was
no surprise since blade erosion is mostly driven by the impact of water droplets
on fast moving bladed. This suggests that being able calculate the amount of
rain that a turbine received along with it blade length and rotation speed, blade
erosion can predicted at 87 percent of the time. For the other 13 percent, the
model of the turbine and the turbines age come into play when the Rainwindlog
alone cannot be used to predict blade erosion

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 16
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

Fig. 17. Model Feature Importance

7 Ethics

Our study attempts to make wind turbine farm operations more efficient and
less expensive, which goes to explain our support for renewable energy. One of
the many benefits of wind energy is that it produces no greenhouse emissions.
However, there are many debates surrounding its benefits. There are concerns
about the environmental and economic effects of wind turbines such as its effect
on landscape, animal habitat and the economy of the towns where they are
built[13].
For wind turbines to be effective they have to be built in high wind areas.
These include mountain tops and ridges. However, people believe turbines built
in these areas ruin the landscape and causes economic loss. A town in Scotland
called Loch Ness where millions of people travel each year to see its hills and
mountains is an example [14]. In 2015 plans were made to develop 30 miles of
wind farms. People protested against the idea claiming that the environment
will lose its natural beauty which will negatively affect the economy of the town.
In addition to that, people believe that this will introduce industrialization in a
place that has been specifically reserved for nature. On the topic of economical
effect some have also argued that besides the large amount of money that needs
to be invested in setting up a wind farm, studies show that Wind Turbines Farms
built near habitats have caused homes to lose their economic value [13].
Presently there is no way of guaranteeing the amount of wind in a given day.
To make up for this unpredictable nature, more wind turbines are built to make
up for low wind days. More wind turbines also mean that much more materials
are needed to produce the turbine. That is, more material per energy produced.
These materials include chemical compound used for making turbine blade, rare
earth metals to create magnets as well as all the steel and concrete needed to
build the turbines. Most of these materials have to be mined from the earth. In

Published by SMU Scholar, 2019 17

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

addition, because of the intermittent nature of wind , power systems that run
on wind energy usually have a backup power supply which is usually fossil fuel.
For this reason Germany a world leader in renewable energy has not been able
cut down its carbon emission significantly even after high investments in wind
and solar energy [15].
Wind Turbine farms have also been known to negatively impact wildlife.
There is evidence of bats and birds colliding with turbine blades according to a
study by the National Wind Coordinating Committee (NWCC)[16]. Even though
the impact is low , the study has evidence of birds and bat deaths from colli-
sions with wind turbines due to changes in air pressure caused by the spinning
turbines. The aerodynamic nature of the blades causes the eardrums of bat to
explode/implode which causes them to lose orientation and fall. The good news
here is that there is ongoing research to reduce bird and bat deaths caused by
Wind Turbines. Researchers have found that bats are most active when wind
speeds are low, so keeping wind turbines off during times of low wind speeds
can possibly reduce bat deaths by more than half without significantly affecting
power production [16].

8 Conclusion

The effect of rain at high blade tip speed as a significant driving force for
blade erosion. This paper confirms this theory and is in agreement with the
scientific consensus on the issue[17]. This work offers a methodology for using
turbine historical fatigue metrics and machine learning to predict blade remain-
ing lifetime. These techniques can be applied to many predictive maintenance
applications in wind turbines and beyond.

9 Future Work

According to the literature[17] there are many particulates that have been
found to contribute to erosion.Data collected on dust, sand, crop debris, and
insects could server to improve model accuracy. There is available data on se-
vere thunderstorms[18] that can help improve the fatigue features and quantify
hail impacts to the blade. Lightning damage is a source of noise for predicting
leading edge erosion.
Since leading-edge erosion can impact the aerodynamics of the blade, there is
an overall performance impact to the turbine. This can be quantified with ma-
chine learning using the first year of life as a baseline. This under-performance
estimation can be used as a feature input to improve model accuracy.

References
1. Wang J. Ni A. J Chen, J. How do wind turbines work?
https://www.energy.gov/eere/wind/how-do-wind-turbines-work, 2019.

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 18
 Martinez et al.: Predicting Wind Turbine Blade Erosion using Machine Learning

2. Michele Froese. Wind-farm owners can now detect leading-edge erosion from
data alone. https://www.windpowerengineering.com/mechanical/blades/wind-
farm-owners-can-now-detect-leading-edge-erosion-from-data-alone/, 2018.
3. VTT Technical Research Centre of Finland. Erosion in wind turbine blades solved
with the help of artificial intelligence. https://phys.org/news/2019-01-erosion-
turbine-blades-artificial-intelligence.html, 2019.
4. M. Jeffrey. Managing turbine operating costs through comprehensive and targeted
rotor blade inspections. Operation and Maintenance Summit.
5. Erik Ela Yih-Huei Wan and Kirsten Orwig. Development of an equivalent wind
plant power-curve. National Renewable Energy Laboratory.
6. Daniel Karlsson. Wind turbine performance monitoring using artificial neural
networks. Department of Energy and Environment.
7. Jose Carlos Moreno Paulo Renato da Costa Mendes Julio Elias Normey-Rico Ce-
sar Hernandez-Hernandez, Francisco Rodriguez and Jose Luis Guzman. The com-
parison study of short-term prediction methods to enhance the model predictive
controller applied to microgrid energy management. 10.
8. L.Yue H.Kim, C. Kim. A study of the growth of single-phase mg0. 8.
9. Jennifer R. Stinebaugh A. Briand D Benjamin A.S. Linsday J. Hill, R. Wind
turbine reliability: a database and analysis approach.
10. S.Moussa D.Fawzy and N.Badr. An approach for erosion and power loss prediction
of wind turbines using big data analytics.
11. Lujing Chen. Basic ensemble learning (random forest, adaboost, gradient boosting)
- step by step explained. https://towardsdatascience.com/basic-ensemble-learning-
random-forest-adaboost-gradient-boosting-step-by-step-explained-95d49d1e2725,
2019.
12. Morde Vishal. Xgboost algorithm:long may she reign! - towards data sci-
ence. https://towardsdatascience.com/https-medium-com-vishalmorde-xgboost-
algorithm-long-she-may-rein-edd9f99be63d, 2019.
13. Cori Lopazanski. Wind energy as an alternative energy source.
https://sites.google.com/site/enviroethenergysources/home, 2012.
14. BBC. Group want end to loch ness and great glen wind farms - bbc news.
https://www.bbc.com/news/uk-scotland-highlands-islands-32286822, 2015.
15. Paul Hockenos. Carbon crossroads: Can germany revive its stalled energy transi-
tion? - yale e360. https://e360.yale.edu/features/carbon-crossroads-can-germany-
revive-its-stalled-energy-transition.
16. Environmental impacts of wind power — union of concerned scientists.
17. Ni AJ Chen, Wang J. How do wind turbines work?
https://www.energy.gov/eere/wind/how-do-wind-turbines-work, 2019.
18. Iem :: Local storm report app. https://mesonet.agron.iastate.edu/lsr/, 2019.

Appendix

Code Repository

The code used for this paper can be found at the GitHub repository below.
https://github.com/CaseyRMartinez/SMU_Capstone

Published by SMU Scholar, 2019 19

 SMU Data Science Review, Vol. 2 [2019], No. 2, Art. 17

Other Resources

https://machinelearningmastery.com/sequence-classification-lstm-recurrent-neural-
networks-python-keras/
https://www.kaggle.com/novinsh/surprise-me-2-with-lstm
https://stats.stackexchange.com/questions/354019/lstm-time-series-classification-
using-keras
https://www.energy.gov/eere/wind/articles/top-10-things-you-didnt-know-about-
wind-power
http://www.windfarmbop.com/category/met-mast/

https://scholar.smu.edu/datasciencereview/vol2/iss2/17 20
View publication stats