1. Introduction
Energy supply chains are fundamental for modern human societies, which require large amounts of energy per capita. In the last centuries, fossil fuels were the main energy source for electricity production, transportation and other economic sectors. The trend, however, is not sustainable because of the environmental footprint and rising exploration costs for oil. To overcome or minimize the limitations of non-renewable energy sources, different renewable sources were proposed, such as biomass, hydropower and wave, geothermal, solar and wind energy. Although energy supply chains have been changing for a renewable reality, there are still many barriers to renewable energy development, such as the conversion cost and efficiency, location selection and distribution network, among others. Wee et al. [
1] report the performance of these new renewable energy supply chains, the existing barriers and how to surpass them.
Wind energy took a key position in the new trend and wind turbines have been widely deployed to generate electric power. Numerous, large and small, wind farms were created in recent decades, in different countries, as part of the global effort to expand production of clean energies from renewable energy sources. Many of the wind farms are offshore, in order to collect higher wind power values and have a lower environmental impact on land usage.
Planning a wind farm is a difficult task. Among other challenges, it is necessary to choose an adequate location. A good location will have good wind power density, stable wind speed, a small environmental footprint and easy access for maintenance, among other requirements.
For optimal performance of a wind farm, it is also important to monitor several variables such as wind speed and temperature of key parts of the wind turbines. Good monitoring and adequate maintenance of wind turbines allow to optimize production and prevent malfunctions that could lead to downtime or even endanger people and property. It can also considerably reduce the maintenance costs of the turbines and support infrastructure. Hau & Erich [
2], and Letcher [
3], offer a comprehensive overview and insight into aspects of wind energy, from historical background to the fundamental science behind the modern industry, covering technical and economic aspects.
Predictive maintenance has been considered to be the best answer for maintenance of wind farms. It allows the extension of components’ lifetime, the maximization of energy output and the reduction of maintenance costs, leading to the performance of corrective maintenance just before equipment failure [
4]. For determining the expected point of failure, it is important to monitor wind turbines and analyze the data collected, using data analysis techniques.
The present paper proposes a survey of the state-of-the-art datasets of wind resources and wind farms. Most of the datasets are available for public use, and they offer a wealth of information which can and must be analyzed for optimal decisions in the process of planning wind farms and optimizing maintenance plans.
The remainder of the paper is organized as follows.
Section 2 presents a literature review.
Section 3 describes some fundamentals about wind as an energy resource and wind turbines’ technology.
Section 4 discusses the characteristics of good datasets.
Section 5,
Section 6 and
Section 7 report important open datasets related to wind turbine capacity and wind farm projects, wind measurements and wind turbine/farm monitoring Supervisory Control and Data Acquisition (SCADA) systems, respectively.
Section 8 gathers other available datasets and a discussion about the existence of data, its quality and comparison of all datasets.
Section 9 presents a deeper overview of a specific dataset.
Section 10 reports the main contributions of the present work to the state of art.
Section 11 draws some conclusions and proposes future work.
2. Literature Review
Numerous research projects analyze existing wind resources and wind turbine monitoring datasets, advancing the state of the art soon after the datasets are available for public use.
González-Aparicio et al. [
5] propose a methodology to capture local geographical information and generate meteorological derived wind power time series, allowing better understanding of the wind resource at wind farms. The study followed up to develop a European wind power generation dataset called European Meteorological derived HIgh RESolution (EMHIRES) [
6]. Both studies mention several sources of wind farm and wind resource databases, such as the Wind Power, Global Wind Atlas and Merra dataset, operational forecast wind speed datasets, European Centre for Medium-Range Weather Forecasts (ECMWF) dataset and wind statistics reports from different countries around the world. Diffendorfer et al. analyze onshore wind turbine locations for the United States [
7], with the purpose of creating a free, centralized, national, turbine-level geospatial dataset, for scientific research, land and resource management. In 2017, the USWTDB (United States Wind Turbine Database) [
8] was created, a national turbine capacity database. Van Vuuren & Vermeulen also report about investigation of wind speed profiles for renewable energy development zones in South Africa [
9].
Predicting the output of a wind farm is an important goal for wind energy industry, and one of the most, if not the most, important variable to look at when deciding to move forward with a wind farm project. Therefore, it is very important to develop performance models. However, it may represent a challenge, since wind turbines’ power is essentially determined by variables which are hard to predict with good accuracy, due to their stochastic nature. Kusiak et al. [
10] examine time series models to predict wind speed and power at different time scales, namely ten minutes and one hour long. They use the wind speed as an input to compute an integrated k-nearest neighbors model, for prediction of wind farm output. The author uses five different algorithms to construct the time series models, in order to select the most suitable for the task. The algorithms used include Support Vector Machine regression algorithm [
11,
12], Multilayer Perceptron [
13], Reduced Error Pruning tree [
14], M5P Tree [
13,
15,
16] and the Bagging Tree [
13,
17,
18]. Research used data generated at a wind farm and collected by a SCADA system, resulting in 4455 recorded instances for wind speed and power at 10 min intervals. The dataset was divided into a 3568 observations training dataset and an 887 observations test dataset.
Computation Fluid Dynamics (CFD) models have also been applied recently [
19]. Most of the models are related to prediction of wind turbines wakes and dynamic behavior, associated with overall power generation. Wu et al. [
20] use Large-Eddy Simulation (LES) to explore the effect of turbine array configurations on the turbine wake characteristics, as well as the power extraction efficiency. The paper associates the impact of the turbines’ hub arrangements and the wind farm power generation. Lin & Porté-Agel [
21] also use LES model to study wind turbine wakes, comparing the prediction results between the two different yaw models. Li & Yang [
22] use the Actuator Disk (AD) model to simulate wind turbine wakes. They also present a study on AD and Actuator Surface (AS) models and their capability to predict dynamic behavior, on utility-scale turbines, for both uniform and turbulent non-uniform conditions. Uchida [
23] also studies the wake characteristics of wind turbines, by predicting them with LES models and parallel computation based on a hybrid LES/actuator line (AL) model. The accuracy from both models is compared and the effects of inflow shear on the wake characteristics is investigated.
When planning a wind farm project, another very important decision is the choice of the turbine with the most suitable characteristics for the place and operating conditions. Pessanha et al. [
24] propose one methodology to analyze anemometer data and evaluate wind potential, in order to help to identify which turbine characteristics should be chosen for maximum profit. The authors use Weibull distribution as a wind speed frequency distribution model at 25 m and 50 m height. Knowing wind speed values at two different heights facilitates estimating wind speed values for other heights, using Equation (
1).
In Equation (
1),
v is the wind velocity desired at height
h,
is the wind velocity measured at height
, and
is the power exponent. After wind power velocity, the authors calculate the average power, from the turbine’s power curve and the Weibull distribution. From average power values it is possible to calculate the Capacity Factor (CF), given by Equation (
2).
Using data from different wind turbine models with different characteristics, the energy production models are computed for each one. The data used are from Sistema de Organização Nacional de Dados Ambientais (SONDA), a Brazilian project to implement infrastructures to survey wind and solar energy resources, with 10 min sampling period, as described in
Section 6.3.
The benefits of monitoring wind turbines are quantified in [
4], where the authors show some of the costs for different maintenance plans. Almost all data collections referenced in the present work have 10 min sampling period. That can be seen as a negative aspect, due to the possible loss of information. Higher frequency data sampling offers more accurate information [
25], although at the cost of using additional computing power. Among other techniques, Principal Component Analysis (PCA) is a useful statistical technique that is applied for data reduction with minimal loss of information [
26,
27]. It is often used in complement with machine learning algorithms [
28,
29].
4. Data Classification and Characteristics of a Good Dataset
The following sections list and describe important wind energy related datasets, which aggregate substantial and important amounts of data. These data are fundamental for modern machine learning applications and Big Data algorithms. There are different definitions of Big Data in the literature. Although some authors focus on the ontological characteristics of the data, others focus on the computational difficulties of processing data. According to Kitchin & McArdle [
31], the concept of Big Data is still being defined, but Big Data datasets must abide by a majority of general traits, such as: (i) Volume (space required to storage data); (ii) velocity (considered a key attribute, it represents the frequency of generation, handling, recording or publishing); (iii) Variety (weakest characteristic attribute); (iv) Exhaustivity (seeking entire population within a system); (v) Resolution; (vi) Relationality; (vii) Extensionality (flexibility of data generation, where a highly flexible data system has a strong extensionality).
The quality of a dataset is directly related to the organization that creates it, and data quality is often related to its value and accuracy. However, data quality has other dimensions, such as uniqueness, completeness, validity and consistency. Also, a good quality dataset must not have errors due to incomplete data, as well as syntactic or semantic errors.
In summary, good open datasets offer a wealth of information that is easily available and should comply with the characteristics referenced above and the following requirements:
Available on the web, in an open format and under a license that permits use for research and other uses;
Available as machine-readable structured data, such as Comma Separated Value (CSV) files or other common data format;
Available in non-proprietary formats, such as CSV or eXtensive Markup Language (XML);
Complies with common open standards and main international standards for the World Wide Web;
Contains enough information about where the data were collected, or link the data to a context;
Contains data points in sufficient quantity and quality for use in data mining, machine learning or other computational methods. The more sensors are monitored the better;
Sampling frequency must be high enough to capture and describe the most important variables;
Ideally there are no gaps in the data, or the gaps are short enough not disrupting the patterns.
The most common definition of data quality is that which determines that the data can fulfill the function for which it was collected.
6. Wind Resource
The wind speed and direction are the most important variables that affect a wind turbine’s output. Hence, accurate predictions and measurements of the wind behavior are fundamental for wind farm planning and management. Among other important decisions, the turbine model must be adequate for the wind available in the place, for maximum efficiency and life cycle.
6.1. OpenEI Dataset
OpenEI dataset [
33] is a trusted source of energy data, specifically for renewable energy and energy efficiency. The information provided is aimed at helping to make informed decisions on energy, market investment and technology development. The data can be viewed, edited and added by the users after the evaluation of content by experts. Open data is part of the core mission for OpenEI and for that purpose, most accessed data on OpenEI comes from several different resources, such as Department of Energy Open Data Catalog (DOE Data), International Utility Rate Database (IURDB) and United States Utility Rate Database (URDB). OpenEI offers information about wind resources, rather than wind farms characteristics. The contents include wind maps, meteorological data and wind power maps, among other variables. In total there are 216 files available in the wind sector. Some of the data used in the present research were extracted from OpenEI.
6.2. Native American Anemometer Loan Program
The Native American Anemometer Loan Program (ALP) was conducted by the U.S. Department of Energy (DOE) and an initiative from Wind Powering America (WPA). The purpose of the ALP was to provide native American tribes a low-cost, low-risk means of quantifying their wind resource, since there were no data to make wind project production and economic performance estimations with precision. The validation process is based on a quality control strategy adopted by the Baseline Surface Radiation Network. By providing native American tribes a low-cost way of quantifying the wind resource in their lands, it was expected that they would be encouraged to pursue wind development, leading to the installation of wind turbines. The program was launched in 2000 and by the end of 2011, 90 towers had been installed over 10 states.
The ALP’s anemometer towers record information about the wind, such as its speed, direction and turbulence, with 10 min sampling periods. All the information was forwarded to the National Renewable Energy Laboratory (NREL) of the USA and analyzed. Free access is given to 11 of the 144 locations that conducted this activity.
Table 2 shows some of the regions with free access to the raw data, available at OpenEI From the regions mentioned in the table, the Navajo Indian Reservation has some missing data, due to malfunction of one of the sensors. Almost all the anemometer towers are 20 m high, because the 20 m high towers were considered adequate for wind turbine projects up to ∼100 kW. All the data of the dataset are provided by NREL. The users are granted the right, without any fee or other cost, to copy, modify, enhance and distribute the data, as long as the users agrees to credit the NREL. The dataset contains monthly average values from more than 700,000 wind observations. The location with less observations available is the Navajo India Reservation, with just 35,661 samples collected in 2004 and 2005. There is a gap in the data, due to a malfunction in the direction vane, which lasted from 18 November 2004 to 22 February 2005. The place with more observations recorded is Northern Cheyenne India Reservation, with a total of 90,891 samples recorded in 2003 and 2004. The Pine Ridge Indian Reservation does not mention how much observations it took to develop the monthly wind speed, direction and turbulence average values.
6.3. SONDA
The SONDA network was born from a Brazilian project to install resources that could track data about wind and solar energy resources in Brazil. Every group of data available passed through a validation process to ensure their reliability, since there are numerous factors that can affect the reliability of the data. Project SONDA [
34] provides wind speed, direction and air temperature at 25 m and 50 m height, with a 10 min sampling frequency. The network has three different stations.
Table 3 summarizes the monitoring period of each survey station of the SONDA dataset. The dataset contains almost six years of good quality data about wind resource on the Brazilian west coast, up to a total of 385,488 observations,
Table 4.
6.4. Ethiopia Wind Measurement Data
Ethiopia Wind Measurement Data is a data repository for measurements collected from 17 wind poles, placed in different regions of Ethiopia [
35]. Data are updated in batches, monthly. Daily wind speed, wind direction, air pressure, relative humidity and temperature reports are recorded.
The wind measurement campaign that generated the data was commissioned by The World Bank with funding from the Energy Sector Management Assistance Program (ESMAP). It is available under The World Bank’s open data policy. Each wind pole contains six different sensors, at different heights, from six to eighty meters, as shown in
Table 5.
A predictive wind resource map can be built using the data collected by the sensors installed at different locations.
Table 6 shows the location of each data collection pole. The sampling frequency is 1 Hz. However, data recorded are the average of 10 min of data samples. Until now there are no known missing data samples. Hence, the available data should be of high quality for mining and studying the variables during the recording period, which in some cases is more than one year and a half.
Sensors installed in different poles may come from different manufacturers. Data for each pole are aggregated in a metadata file that also contains information about the pole and the sensors’ manufacturer, model and serial number.
6.5. Global Wind Atlas
The Global Wind Atlas (GWA) is a free, web-based application developed as an aid for policymakers, planners, and investors, to help identify high-wind areas for wind power generation, possibly anywhere in the world [
36]. It is available at
https://globalwindatlas.info (accessed on 17 August 2020).
GWA provides an online service where users can search through different queries. It also provides free downloadable datasets, grouped by different sections, based on the latest input data and modeling methodologies.
Table 7 shows a summary of the downloadable sections in the Global Wind Atlas.
In the GWA it is also possible to download high-resolution maps of the wind resource potential at a global and country level for 10 m to 200 m height. This wind resource database is maintained in partnership by the Department of Wind Energy at the Technical University of Denmark and the World Bank group. The mesoscale model mentioned on
Table 7 uses ECMWF ERA-5 reanalysis data for atmospheric sampling for the period 1998–2017. ERA5 provides hourly estimates of many atmospheric, land and oceanic climate variables. The data cover the Earth on a 70 km grid and resolve the atmosphere using 137 levels from the surface up to a height of 80 km. ERA5 includes information about uncertainties for all variables at reduced spatial and temporal resolutions.
The output at 3 km resolution is generalized and downscaled further using WAsP software plus terrain elevation data at 150 m resolution, and roughness data at 300 m resolution. The WAsP software suite is the industry-standard for wind resource assessment, siting and energy yield calculation for wind turbines and wind farms. Finally, the microscale is sampled on calculation nodes every 150 m. However, this modeling process becomes more uncertain, most likely leading to an overestimation of mean wind power values.
Table 8 shows the spatial scales for different length scales of the GWA, according to [
3].
7. Wind Farm Monitoring
The uncertainty on revenue of the existing wind farm installations pressures operation and maintenance departments to reduce their costs, since they might come up to 30% in offshore environments [
4]. Due to factors like the ones mentioned, monitoring systems focused on wind farms’ behavior and main components have been increasingly installed to optimize maintenance planning. Their economic benefit has been investigated and proven to exist [
4].
7.1. ENGIE, La Houte Bourne Wind Farm
ENGIE is a company that produces and distributes energy from different sources, including renewable sources. It is a major player in the production of green electricity, being the 1st wind power producer with installed capacity of 1730 MW, representing approximately 8% of France wind power production, according to power production in
Table 9.
According to the company’s strategy, ENGIE decided to open up, for public use, data of the La Houte Borne wind farm, which is operated by ENGIE Green. The farm’s four wind turbines are all from the same model and manufacturer, and have been providing electricity to the equivalent of 7300 people since 2009, avoiding 12,000 metric tons of CO
emission per year. The data set is composed by two very big and complete spreadsheet files, one from 2013 to 2016 and another from 2017 to 2020. The files contain information about each wind turbine’s components, such as rotor speed, nacelle temperature, mechanical information like the torque and, finally, wind speed, direction and pitch angle. All the variables are cataloged and described in another file named “Data Descriptions”. Finally, ENGIE provides the static information about each turbine: ID number, manufacturer, model, rated power, hub height, rotor diameter and precise location, as described in
Table 10. All the data are recorded with a ten-minute period, and there is a total of 1,057,868 observations.
Data are in a file with SCADA data about component control variables and meteorological mast. Data description is provided, with every variable abbreviation explained and units. Static information contains some of the turbine’s technical characteristics.
7.2. Sotavento Wind Farm
Sotavento wind farm [
38], was put into operation in 2001 by Sotavento Galicia, S.A, in Xermade, Lugo, Spain, after the Galician Government had decided to increase investment in renewable energy, especially wind-based energy. Composed by 24 onshore turbines with a total nominal power of 17,560 kW it has an average annual generation of 33 MWh and produces the equivalent consumption to 1051 families, avoiding 0.36 MT of CO
emissions per hour and consumption of 0.68 barrels of petroleum.
Sotavento’s platform is a reliable source of wind resource and wind farm output production. With a high-quality monitoring program, the database provides real time data for wind speed and direction, turbines’ production and capacity factor, and finally temperature and density of the air. All the information is given with a 10 s period. Additionally, it is possible to search for historical data, allowing researchers to study wind farm and meteorological mast behavior.
Table 11 shows a summary of the turbines in use in the wind farm. The historical data is logged with 10 min intervals, hourly and daily.
7.3. EDP Wind Farm
EDP (Energias de Portugal) is an important player in the energy sector, especially in the Iberian Peninsula, where it produces and distributes a large share of electricity. The dataset available provides two years of SCADA records from five offshore wind turbines located in the West African Gulf of Guinea [
39]. The dataset consists of different files that give information about failure logs and technical information about some of the main turbine’s components, such as the gearbox, generator and rotor. Additional information includes meteorological data, namely wind speed and direction, air pressure, humidity, temperature and component signals, namely generator RPMs and oil temperature in the hydraulic group. All files available are summarized in
Table 12, including listing of all variables logged. The training set is from 2016 (all year) and the testing set consists of nine months of data, from 2017 (1 January 2017 to 1 September 2017). Meteorological mast data and component signals are recorded with a 10 min period and there is a total of 69,962 observations.
The Wind Turbine Characteristics file contains wind turbine main characteristics,
Table 13. Also, it supplies wind turbine’s power curve, a defining variable, at a 1.225 kg/m
air density.
The meteorological mast file logs important meteorological signals, namely: Anemometer sensors 1 and 2 are at a 80 m and 77 m height; Weather vanes are located at 77 m and 40 m height; and temperature and pressure sensors at 75 m and 100 m height.
The component signals file includes SCADA signals for each wind turbine’s most important components and production values.
The failure logs file is an historical failure logbook for the wind farm. It logs replacement and repaired processes, errors, high signal values and component failures.
7.4. Yalova Wind Turbine Dataset
Yalova is an onshore Turkish wind farm, located in west Turkey. It comprises 36 wind turbines and a total nominal power of 54,000 kW, with two different turbine models, according to
http://www.tureb.com.tr/bilgi-bankasi/turkiye-res-durumu (accessed on 18 May 2020).
Table 14 summarizes the turbine models and characteristics.
The Yalova wind farm has been operating since 2016. A SCADA system was used to measure and save wind turbine’s data from one of the turbines—which model is monitored is not specified in the dataset. The SCADA system logged wind speed and direction, generated power and the theoretical power based on the turbine’s power curve. Each new line of data is stored at 10 min intervals. However, there are a few gaps and some generated power is missing, which can be explained as a wind turbine’s malfunction, maintenance or the wind speed being lower than the cut-in speed. The dataset is available in CSV format and it is for a one-year period, at
https://www.kaggle.com/berkerisen/wind-turbine-scada-dataset (accessed on 17 August 2020). All the information about the data available is summarized on
Table 15.
7.5. Wind Turbine SCADA dataset
In Kaggle website there is a data file, downloadable at
https://www.kaggle.com/wasuratme96/turbine-fault-prediction (accessed on 17 August 2020). Kaggle is one of the world’s largest data science communities. One of the community members released a massive SCADA dataset from an unknown turbine. The turbine’s location or identification were not disclosed, but the data set seems to provide a wealth of information, including wind speed, power production, operating hours, component monitoring and different turbine status.
The data are divided in two files: one with wind and turbine’s behavior and the other with turbine status, including status “under maintenance”. The former file contains logs of wind forecast, turbine’s generating power, component’s temperature and available power in the wind. The latter file contains logs of maintenance periods and failures.
Table 16 shows some details of the files available in the dataset.
9. A Deeper Overview of EDP Dataset
As mentioned in
Section 7.3, EDP dataset is one of the most complete datasets available. Among other information, it is possible to search for relations between turbine failures and their behavior days before the failure.
9.1. Wind Turbine Operation Behavior
Wind Turbines have three main regions of operation, [
40], as shown in
Figure 1. The regions are:
Region 1: Includes the time when the turbine is starting up;
Region 2: Operational region in which it is desirable to seize as much wind power as possible;
Region 3: Wind speeds are relatively high (rated wind speed) and force the turbine to limit the fraction of wind power captured, for electrical and mechanical safety.
Table 20 shows the number of failures counted in the EDP dataset, grouped by turbine component group and sorted in descending order by number of failures counted. In the table it is possible to identify three groups of components more often affected by failures. They are: (i) Generator; (ii) Generator Bearing; and (iii) Hydraulic Group.
Figure 2 shows that Turbine 06 and Turbine 09 are the two most affected turbines—the former counts seven failures, the latter counts five failures.
Figure 3a,b show plots that help identifying in which region the turbines were operating before failure. Prior to failure in turbines T06 and T11, Region 2, the desirable operational region of operation, is also identified as the one more prone to failure. Despite the fact that turbine T06 shows some prior to failure behavior near Region 3 of operation, most of the observations still tend to Region 2.
9.2. Principal Component Analysis
PCA is a statistical method to identify patterns in data and express them in a way to highlight the similarities and differences, through a graphical representation. PCA is also a method to compress the size of the datasets, reducing the noise, removing outliers and simplifying data description.
Table 21 shows the correlations between different variables of the observations. The variables are: (i) Generator’s rotations per minute; (ii) Generator’s bear temperature; (iii) Hydraulic oil temperature; (iv) Gear oil temperature; (v) Nacelle temperature; (vi) Rotor’s rotations per minute; (vii) Wind speed; (viii) Ambient temperature.
As the table shows, some of the variables have high percentage of correlation, which means their behavior is highly correlated. For example, the Generator RPM shows very high correlations () with Rotor RPM, which is easily understood, since the behavior of one component is directly connected with the other. Nacelle temperature has high correlation with other temperature-related variables, less with the hydraulic oil temperature. Ambient temperature does not really correlate with any variable chosen. We might have thought otherwise, but this case shows no behavior correlation.
Since there are more than 200,000 observations, it is not possible to plot the raw SCADA records. An alternative is to compress the data to a lower number of dimensions, based on PCA approaches.
The generator component is the one with more failures recorded. Hence, it was chosen for this PCA implementation. Nine variables with high correlation levels were selected, from two distinct turbines, creating a 9-dimension plot. Using PCA it was possible to reduce the observations-variables plotting to only a 2-dimensional chart, without losing relevant levels of information.
Figure 4b and
Figure 5b show that by only using the first two principal components, it is still possible to have over 80% of variance explained. PC plots in
Figure 4a and
Figure 5a show that zero, one and two days before failure, turbine’s behavior is clustered with approximately eight, nine and ten days before failure. This observation might indicate that it may be possible to predict a high probability of failure 10 days before it occurs.
levels are often used for control charts, which give a statistical measure of the multivariate distance between each observation from the center of the dataset.
Figure 6 shows the plot of
for both turbines. As the plots show, T06 has a less erratic behavior and, as a first reading and impression, the process looks under control, with only two out of control moments. T11 has a more scattered
plot, but still it is possible to identify some peaks that might surpass control chart’s limits.
Although we do not have values for control charts, by looking at
plots,
Figure 6, we see similarities too and we can connect them it the PCA plot. The statistical measure has high peaks in the first observations (ten, nine and eight days before failure) and further close to failure.
11. Conclusions
Wind energy has been the main renewable energy source in recent decades and it has potential to continue growing. Good quality open datasets of wind resources, wind farms and wind farm operation are fundamental for researchers, to extract knowledge and advance future research. The present paper proposes, therefore, a comprehensive survey of existing datasets, with their advantages and limitations. A total of 15 open datasets were analyzed, 13 of which have good quality for machine learning applications. The main characteristics of good quality datasets have been pointed out. This is an important contribution to facilitate future research in the field.
A deeper analysis of one of the most complete wind farm operation datasets available also provides these conclusions:
The performance of component groups differs, and the faultiest behavior was identified;
Wind turbine’s region 2 of operation is where more failures occur, even though it requires less mechanical effort;
Using PCA it may be possible to predict a turbine failure up to 10 days before the actual failure.
Future work includes the use of the knowledge extracted from the datasets to improve turbine maintenance plans, so that the number of failures, downtime and corrective maintenance costs may be reduced.