STOCKHOLM SVERIGE 2017,
Automatic wind turbine operation analysis through neural networks
Automatisk driftanalys av vindturbiner medels neurala nätverk
ALEXANDER BOLEY
KTH
SKOLAN FÖR ELEKTRO- OCH SYSTEMTEKNIK
www.kth.se
School of Electrical Engineering
Automatic wind turbine operation analysis through neural networks
Automatisk driftanalys av vindturbiner medels neurala nätverk
Alexander Boley
Degree Project in Electrotechnical Theory and Design, Second Cycle
EI270X, 30.0 credits
Author: Alexander Boley Supervisor OX2: Olle Viotti Supervisor KTH: Per Westerlund Examiner: Patrik Hilber
Abstract
This master thesis handles the development of an automatic benchmarking program for wind turbines and the thesis works as the theoretical basis for this program. The program is created at the request of the power company OX2 who wanted this potential to be investigated. The mission given by the company is to: 1. to find a good key point indicator for the efficiency of a wind turbine, 2. to find an efficient way to assess this and 3. to write a program that does this automatically and continuously. The thesis determines with a study of previous research that the best method to utilize for these kinds of continuous analyses are artificial neural networks which can train themselves on historical data and then assess if the wind turbine is working better or worse than it should with regards to its history. This comparison between the neural network predicted operation and the actual operation works as the measurement of the efficiency, the key point indicator for how the turbine work compared to how it historically should operate. The program is based on this principle and is completely written in MATLAB. Further testing of the program found that the best variables to use are wind speed and the blade pitch angle as input variables for the neural network and active power as the target used as the variable to predict and assess the operation. The final program was able to be fully automated and integrated into the OX2 system thanks to the possibility to continuously import wind turbine data through APIs. In the final testing was the program able to identify 75% of the anomalies manually found in the half year and in the five turbines used for this thesis, the small anomalies not found manually but identified by the program excluded.
Key words:
SCADA, artificial neural network, wind turbine, data analysis, statistical analysis, O&M, MATLAB
Sammanfattning
Den här masteruppsatsen hanterar utvecklandet av ett automatiskt driftanalyseringsprogram för vindkraftverk och fungerar som det teoretiska underlaget för detta program. Programmet utvecklades på uppdrag av kraftbolaget OX2 som ville undersöka potentialen för ett sådant analysprogram i deras verksamhet. Uppdraget givet var att: 1. ta fram en bra indikator när det gäller den faktiska effektiviteten av ett vindkraftverk, 2. att hitta ett effektivt sätt att använda detta måttet i en analys där målet är att hitta avvikelser, och 3. skriva ett program som automatiskt kan använda måttet och metoden över tiden. Rapporten kommer via litteraturstudie fram till att tidigare forskning visar på att neurala nätverk är den mest lovande metoden för att genomföra sådan här analys. Dessa nätverk kan träna sig själva på historiska data och sedan analysera om vindturbinen arbetar bättre eller sämre än historiskt. Den här jämförelsen mellan den historiskt grundade förutspådda kraften ut och den faktiska kraften ut fungerar som kvalitetsmåttet på hur bra turbinen fungerar.
Programmet är baserat på den här principen och är helt skriven i MATLAB. Vidare tester av programmet visar att de bästa variablerna att använda för att förutspå kraften ut är vindhastigheten och bladens vinkel mot vinden.
Slutprogrammet var kapabelt att fullt automatiskt och integrerat i OX2s system identifiera 75% av alla avvikelser som manuellt hittats i ett halvårs data på de fem turbinerna använda för rapporten, småfel hittade av programmet men inte manuellt exkluderat.
Nyckelord:
SCADA, artificiella nätverk, vindkraft, dataanalys, statistisk analys, drift och underhåll, MATLAB
Acknowledgements
A chest full of thanks to OX2 and Olle Viotti who not only made this thesis possible but also gave me an office space and full access to their data to allow me to make the most of this opportunity. This thesis would not have been able without the continuous support and feedback from the colleges at OX2 throughout the whole writing of the thesis.
I also want to thank Per Westerlund who always were ready to safely guide me throughout the writing, the theory and the labyrinth-like administrative process of KTH.
Table of contents
Abstract ... ii
Sammanfattning ... ii
Acknowledgements ... ii
Table of contents ...iii
List of figures ... v
List of equations ... v
List of tables ... v
Abbreviations ... vi
1. Introduction ... 1
1.1. OX2 ... 2
1.2. Aim and objective ... 2
1.3. Thesis boundaries ... 2
1.4. Company secrets ... 2
1.5. Thesis Overview ... 2
2. Literature Review... 3
2.1. Wind energy dependencies ... 3
2.2. SCADA data... 3
2.3. Analytical and statistical methods ... 4
2.3.1. Fuzzy inference system ... 4
2.3.2. Artificial neural network ... 4
2.3.3. Adaptive neuro-fuzzy inference system .... 4
2.3.4. Signal trending ... 4
2.3.5. Physical models... 4
2.4. Similar works ... 5
2.5. Literature review conclusion ... 5
3. OX2 and Wind Power... 5
3.1. Wind turbines ... 5
3.2. The chosen wind turbine field ... 6
3.3. Wind turbine type ... 6
3.4. Variables ... 6
3.4.1. Wind speed ... 6
3.4.2. Ambient temperature ... 6
3.4.3. Wind direction... 7
3.4.4. Nacelle direction ... 7
3.4.5. Air pressure... 7
3.4.6. Active power ... 7
3.4.7. RPM ... 7
3.4.8. Pitch angles ... 7
3.4.9. Operational state ... 7
3.4.10. Data availability and IEC-availability ... 7
3.5. Data ... 7
4. Methodology... 8
4.1. Basic calculations... 8
4.1.1. Errors... 8
4.1.2. Standard deviation ... 8
4.1.3. Clusterification ... 9
4.1.4. Mahalanobis distance ... 9
4.2. MATLAB ... 9
4.2.1. clusterdata ... 9
4.2.2. corrcoef... 9
4.2.3. NN training ... 9
4.2.4. mahal ... 9
4.2.5. listdlg... 10
4.2.6. Plots ... 10
4.2.7. datetime ... 10
4.2.8. tic, toc ... 10
4.3. Logical structure ... 10
4.3.1. Function and save based program... 10
4.3.2. Options ... 10
4.3.3. API ... 10
4.3.4. Preparation of data ... 10
4.3.5. Neural Network training ... 11
4.3.6. Error estimation ... 11
4.3.7. Training rerun... 12
4.3.8. Automation ... 12
4.4. Setup of program testing ... 12
4.5. Result analysis ... 12
4.6. Boundaries... 13
5. Results... 13
5.1. Program optimization and design ... 13
5.1.1. Clusterification ... 13
5.1.2. Impact of a second neural network run ... 14
5.1.3. Time step length ... 14
5.1.4. Neural network training theory ... 14
5.1.5. Single or several targets ... 15
5.1.6. Neuron layers ... 15
5.1.7. STD for anomaly classification ... 15
5.1.8. Filter choice ... 15
5.1.9. Variable choice... 16
5.2. Program final appearance and result summary 17 5.3. Automatic program final structure ... 17
5.4. Final anomaly search... 20
6. Discussion ... 20
6.1. Sensitivity analysis... 21
6.2. Future work ... 21
7. Conclusion... 22
8. References ... 23
List of figures
Figure 1. - Power supply of the World 2015 [3] ... 1
Figure 2. - Power supply of Sweden 2015 [6]... 1
Figure 3. - Swedish Wind Power Supply 2015 [8] ... 1
Figure 4 - Standard wind turbine design ... 6
Figure 5 - Confusion matrix ... 12
Figure 6 - Clusterification MAE impact ... 13
Figure 7 - Clusterification TPA and AA impact ... 14
Figure 8 - Second run MAE impact... 14
Figure 9 - Second run TA and AA impact... 14
Figure 10 - Time step length MAE impact ... 14
Figure 11 - Training theory MAE impact... 14
Figure 12 - Training theory anomaly acc. ... 15
Figure 13 - No. of layers impact on AA... 15
Figure 14 - STD for anomaly classification ... 15
Figure 15 - Filter impact on MAE ... 15
Figure 16 - Percent of the data removed ... 16
Figure 17 - True positive acc. ... 16
Figure 18 - Anomaly acc. ... 16
Figure 19 - Filter comparison graph ... 16
Figure 20 - Variable choice MAE impact ... 17
Figure 21 - Variable choice AA impact ... 17
Figure 22 - Program Structure ... 19
Figure 23 - Concept Graph ... 20
List of equations
Equation 1 - Absolute error. ... 8Equation 2 - MAE ... 8
Equation 3 - MSE ... 8
Equation 4 - STD... 8
Equation 5 - TA... 13
Equation 6 - AA ... 13
List of tables
Table 1 - Program design choices ... 17Table 2 - Final anomaly search ... 20
Abbreviations
AA – Anomaly accuracy
ANFIS - Adaptive Neuro-Fuzzy Inference System ANN - Artificial neural network
AP – Air pressure AT – Air temperature CM - Condition Monitoring
CMS - Condition Monitoring System FDS - Fault Detection System KPI – key point indicator MAE – Mean absolute error ND – Nacelle direction NN - neural network
O&M - operation and maintenance PI – Pitch angle
RS – Rotor speed
SCADA - Supervisory Control And Data Acquisition TPA – True positive accuracy
WD – wind direction
WEC - Wind Energy Converter WS – Wind speed
WT – Wind turbine
1. Introduction
The worlds’ energy system is undergoing one of the largest changes since the start of using oil. This is due to the fact that the international community has recognised the problem posed by manmade global warming and are now trying to take steps to avoid a global catastrophe [1].
A large part of this change is around the way electricity is produced as fossil fuels are used extensively in the power production sector as seen in Figure 1. Energy efficiency is one way to avoid this as a way to decrease the electricity consumption but it will not be enough to stop the use of fossil fuels. To change the production methods are also required and both companies and governments are working hard to create renewable electricity generation that are able to compete economically with the fossil industry [2]
Figure 1. - Power supply of the World 2015 [3]
The electricity market in Sweden faces great challenges even though the actual fossil based power production is very low as seen in Figure 2. This is due to the large part nuclear energy that is going away in the coming years.
The nuclear reactors in Sweden are old and will soon have to be phased out and the government of Sweden have chosen not to invest in new nuclear power plants and instead favours a vision of an electricity system with 100% renewable energy. This is realistic thanks to the hydro power already installed and the biomass potential of the country but also thanks to a fast-growing wind power industry [4] [5].
Figure 2. - Power supply of Sweden 2015 [6]
In Sweden, wind power has been the main focus in the changing of the power production and is highly favoured over for example solar due to the regional conditions. The wind’s potential is quite stable most of the year while the solar irradiance is close to negligible in the whole winter season [7]. The potential of wind power, the opportunities in the changing market and the support from the
government are all important factors and explanations for how the Swedish wind power market has grown
exponentially the last 20 years as seen in Figure 3.
Figure 3. - Swedish Wind Power Supply 2015 [8]
In contrast to the extensively developed power production methods like thermal and hydro, wind power is still quite young and under development. Every year progress is made with new operational methods and technical
solutions that improves the systems. The first and old wind power plants that are now being taken down are a world away from the ones that is being built today even though they are quite young compared to other power plants in operation [9]. But this development is not only seen with wind power but in all renewable power production methods that are entering the market and these renewable technologies are fighting for the economical superiority that will make them the investment of the future [10].
Despite the growing market and the increased efficiency of wind power, the Swedish electricity market are making it hard for all power producers to be profitable as the price of electricity is lower than the cost of producing it [11]
Coal 41%
Oil 4%
Natural gas 22%
Nuclear 11%
Hydro 16%
Other 6%
Hydro 41%
Nuclear 30%
CHP 12%
Wind 9%
Import 5%
Other 3%
0 5000 10000 15000 20000
1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015
GWh to grid
Year
[12]. For the wind power market to thrive despite the low electricity price and the heavy competition the profitability has to be raised [10]. One way of doing this is increasing the operation efficiency of the wind turbines. The majority of the wind turbines in Sweden are connected to a SCADA (Supervisory Control And Data Acquisition) database were data from a myriad of sensors are collected and stored. Through deeper analysis of this data, trends and faults could be found with efficient and automatic processes to make the operation of the turbines more efficient and thereby increase the profitability of and competitiveness of the turbines [13].
1.1. OX2
OX2 is a growing Swedish company that focuses on developing, building, financing and managing renewable power plants in the Nordic countries. These power plants consist mainly of wind power but they are also having projects in solar, geo thermal and bio energy fields [14].
Their wind turbines are connected to an SCADA database and they recognize a larger potential of statistical data analysis in this database and are trying to develop more methods to increase the efficiency of their wind power operation and maintenance. OX2 focuses on a slim organization and thereby automatization of the
surveillance of their many power plants. This thesis is a part of this continuous work and will focus on finding methods that could be used to automatize the process of finding anomalies in the power production through analysis of the incoming SCADA data.
1.2. Aim and objective
This thesis has been conducted at the company OX2 with the overall aim to make their wind turbine operation more efficient through automatic analytical processes of their SCADA data. This will be done by testing different statistical methods and compare variables to find which methods and variables gives the most useful results to the company. The process will then be made automatic and continuous with a program interconnected with the OX2 database that analyses the data and sends an alarm with if it detects strange behaviour in the interconnection of the variables. The main objectives are:
- Find KPIs that are good indicators on the wind turbines performance.
- Find which methods are most efficient in finding hidden efficiency deviations in the data.
- Create a program that can find anomalies automatically.
1.3. Thesis boundaries
The foundations of the methods used can be used for very different purposes. As an example, similar methods used
in this report could be used to find errors in the components and can be used to better schedule maintenance before the braking points of these. This requires extensive research in the individual component behaviour before and after faults and reparations and will be outside the boundaries of this project, due to time limits, but could be a future continuation of this thesis that will focus only on finding anomalies during normal operation. The program produced during this thesis will not result in a completely finished product. It should work and produce useable results but the bug testing and overall implementation of all OX2 turbines will fall outside the limits of the work.
1.4. Company secrets
OX2 has a strict secrecy policy as they handle a lot of companies’ information due to their intermediary role. As this thesis are initiated on their request and are completely reliant on their and their customers’ data, almost all actual data and results are kept secret and will not be presented in this report. This includes the turbine models, the owner of the turbines, the positioning of the turbines, the size of the turbines and the wind farm. Instead other means of presenting and comparing the results will be used
throughout the report to still be able to present the results, these are explained in detail in the methodology section.
The program written during this thesis will end up being OX2 property and will thereby also be included in the company secrets and will as such not be presented even though the logic and functions will be explained
throughout the thesis.
1.5. Thesis Overview
The report will follow a standard report structure and start with an introduction that contains overall information of the thesis role in the larger picture and the basis of the research questions asked. Then a literature review will be presented with an overview of the earlier research in the area and which methods seems to work. Then the methodology will present the methods chosen based on the questions asked and the basis of the literature review.
The methodology will also contain the function used during the programming and the overall logic this is based on. The results produced by the program will follow and present which methods worked best at finding anomalies and how these are implemented in the program. The final programming logic will also be presented at the end of this section. Then the discussion section will argue over the result and what can be made different while the
conclusion after this summarizes the overall thesis and its results.
2. Literature Review
2.1. Wind energy dependencies
The power in the wind can be calculated by using the standard kinetic energy formula seen below as Equation 2.1.
-EJAPE?AJANCUL I R6
t Eq 2.1
Where m is the mass and v is the speed. In the case of wind turbines, the speed is the wind speed and the mass is the mass of the air moving through the rotor. To
understand the kinetic energy equation in regards to wind energy the mass can be expanded as in Equation 2.2.
9EJ@ AJANCUL Ø # R R6
t Eq 2.2
Where Ø is the density of the air and A is the area covered by the wind turbine rotor. But just like the mass, the density is depending on several variables relevant for wind turbines including air temperature and air pressure which are changing depending on things like the turbulence and the weather. The density can be estimated with the ideal gas law if these variables are measured on the wind turbine. As seen in equation 2.2, the most important variable is the wind speed as its value are increased to the power of three when calculating the wind energy. The second most important variable for increased wind energy are the radius of the rotor as the length of the wings increases the area covered exponentially. The density of the air can not be changed in other ways than placing the wind turbines where the environment produces high densities but such changes are marginal. By approximating these variables, wind speed, wing length, air temperature and air pressure the energy in the wind can be
approximated. But all this energy can not be converted to electricity by a wind turbine. This is due to the wind behind the turbine being slower than the incoming wind due to the energy loss which leads to the incoming wind having to push the slower wind away to reach the turbine and thereby losing some of its kinetic energy. This theoretical limit of how much energy can be converted from the wind are called Betz’s limit and are
approximated to 59.26%. This limit is theoretical and can not be reached. The actual limit is depending on the circumstances and the specific turbine [15].
2.2.SCADA data
Supervisory Control And Data Acquisition (SCADA) is an umbrella term for a method of networking different
systems to compile and centralize data to a central supervisory control system. SCADA is used in various processes to oversee, monitor and control advanced and
varied systems. The ability to control and supervise different separate systems in a central interface makes the operation of the associated processes manageable. These characteristics makes it suitable for wind turbines. The turbines have many components and sub processes all with their own measuring points as the steps from kinetic energy in the wind to electric power in the electrical grid requires several transformations, each with different sub systems [16] [17] [18].
The benefits of using SCADA are numerous on its own from a supervisor’s point of view but the compiling of several data systems has laid a foundation for completely new condition monitoring systems (CMS) that can use the gathered data from several separate systems and find correlations that earlier where hidden [19]. The CMS can be used to find fault in both the operation of the turbine and the different components. This could advantageously be used to find error in the operation that the separate systems or software doesn’t detect or to predict when a component is being stressed and thereby in need of repairs [17].
One of the SCADA data analysis strengths is that it doesn’t require additional sensors or equipment. Rather the already collected sensory data is evaluated to get efficiency increases. This makes it a very profitable field to delve into for wind power companies [13].
There are several problems that can arise in the day to day operation of a wind turbine that does not have to be affected by components braking or similar problems.
Instead the problem can be based in software errors, weather phenomenas or similar unforeseen things. As these are not planned for they seldom produce an error alarm to the operator and are thereby hard to find. An automated CMS could find and identify these errors in the SCADA database and thereby increase the efficiency of the system [13] [20]. One such an example is a OX2 case where the wind turbine dropped 30% of its output, seemingly without reason. No alarms or systems were activated to warn about this behaviour, which returned for several weeks before it was noticed. A simple automated SCADA analysis program could have noticed this behaviour within the first hour and the problem would have been detected and solved in its cradle instead of after several weeks. [21]
To use the data for finding faulty components other similar methods are used. The most common one are vibration sensors but other methods include thermography, ultrasonic sound, radiography, oil analysis and more [22].
These methods are used to find components that are
starting to brake and these are then repaired or replaced accordingly. But components can take time to order and wind turbines are often placed far off which creates long logistic chains. Thereby it is important to find faults early to be able to plan and perform the maintenance required as efficiently as possible to maximize the availability of the turbine. CMS of the SCADA data could be an additional tool in this work as it could be used to create new possibilities to find faulty components earlier than the common sensors [16] [23]. This has been shown to work well in a number of reports that have tested different methods of prediction. Notably: [24] [25] [26] [27] [28].
2.3.Analytical and statistical methods
The statistical methods used to analyse SCADA data related to wind turbines are numerous even though the field is quite new [16]. Furthermore, the ways they are used and implemented are even more varied but in this section the most common will be presented with their respective methods. None of the methods are completely unattached and they can all be, and are, used together with the intention of creating better systems. Several reports are presented in section 2.5. Similar works to delve deeper into these combinations.
2.3.1. Fuzzy inference system
Fuzzy logic is exactly what it sounds like, a logic where variables coexist with a degree of vagueness rather than true or false. It is a logical method that can be very useful when evaluating large highly complex systems where the behaviour of the variables is hard to understand and predict. Another useful property is that experience can be included in the fuzzy calculations by tuning the
membership functions used and thereby get the results close to reality. The fuzzy logic works well with wind turbine data and have a good ability to find patterns where the variables aren’t consistent [26] [17] [13] [29].
2.3.2. Artificial neural network
An artificial neural network (ANN) is essentially a learning algorithm where the ANN learns a specified pattern through creating different kinds of interconnected neurons. These are then organized in one or multiple layers. The neurons can have several different functions, for example summation of all data it gets as input from other neurons or it doesn’t give an output before the average input values are positive. The creation process of such an ANN is based on the biological neural network in the brain and rather than programming this complex set of neurons by hand the program is being trained on these patterns by testing which combination of neurons and connections gets closest to the wanted result. When a pre- set error margin is reached, the training stops and the
ANN is ready to be used on similar problems it has been trained on, like for example recognize if its a human face on a picture or not. There are several different training methods used to guide the program in creating its ANN, these are depending on what the goal of the ANN would be as finding patterns in pictures are different to finding patterns between variables [16] [30].
2.3.3. Adaptive neuro-fuzzy inference system The adaptive neuro-fuzzy inference system (ANFIS) combines the logic of fuzzy interference systems with the artificial neural network method to create a new method that combines the strength of both theories to create a program that works well with realistic problems in non- ideal situations. The FIS is used as an early training layer to steer the ANN in the right direction depending on fuzzy logic that steers the program to a solution more attuned to reality. This early steering is also helping the training process to reach good results faster and are thereby very advantageous for large systems [17] [31].
2.3.4. Signal trending
The signal trending method uses statistical data that should be similar to try to find differences. When evaluating wind turbines, the data used could be from the same turbine in an earlier time or different similar wind turbines,
preferably in the same location. This statistical data is compared to find anomalies as different data in the same situations can be a good indicator for fault prediction. For example, if the gearbox oil temperature suddenly starts to spike repeatedly compared to a nearby turbine, it is probably something that should be looked at. Often this method is used to create trend curves which are used to warn when the data deviates [16] [32].
2.3.5. Physical models
The physical modelling approach requires a good
understanding of the system being analysed as it is based on the physicality of the components rather than statistics or comparisons. The method uses physics formulas to calculate the standard behaviour and the interaction between variables and then uses this to assess if for example a turbine works accordingly. The downside with these kinds of calculation could be the amount of work required to fit the physical functions to the actual data as they often are based on laboratory experiments with isolated situations. When applied, variables could be needed but that have not the required sensors installed on the turbine or if installed, have data that fluctuates to much depending on for example turbulence. This requires work with fitting the formulas according to the specific wind turbine data [33] [16]
2.4. Similar works
In this section, several reports will be presented that solves relevant problems with interesting combinations of methods to try to increase the efficiency of the systems.
Zaher et al focuses on a ANN that evaluates the SCADA data in search of temperature deviations for different components through their oil sensors. They identify the usefulness of this as its possible to early identify faulty components and sort through SCADA data in the search for component anomalies [24].
Sun et al uses fuzzy logic to compare different data sets and ANN as a complement to sift through the result from the fuzzy system with a methodology close to ANFIS. The three data sets used are 2 months from the current WT, 2 months from the current WT but a year earlier and 2 months from a WT nearby. The concluded that the prediction models based on these three sets differs noteworthy from each other but can still be used better than old system with boundary condition for the different variables [13].
Schlechtingen et al compares several methods including regression analysis and ANN to find a number of known errors in the wind turbines. They conclude that the regression models strength is its simplicity and even though it finds the faults, its completely outclassed by the ANN as it can not take into account things like seasonal effect which leads to the ANN finding the same faults but earlier in the process when they still are small [34].
Chen et al tries to create a superior fault detection ANN with the help FIS and are thereby using ANFIS methods to find errors foremost in the pitch control system. The report focuses on comparing WT data before and after
maintenance and classifies these as “good” or “bad” to create a computer model that can find early indicators of the bad behaviour and thereby help the operators to schedule maintenance accordingly [26].
Bi et al are investigating if normal behaviour models can be used together with ANN to find wing pitch errors earlier than the systems used today and concludes that indeed its possible. In the six cases tested the prediction could be improved by 13 to 20 hours compared to the commonly used alarm systems [27].
Saleh et al delves deeper into the fuzzy calculation
methods and how these should be designed and connected to an adaptive network rather than a trained ANN. The report come to the same section as presented in the
previous section, that using the more rigid statistical methods rather than ANN creates problem when seasonal changes are starting to affect the system. The hybrid system presented solves this problem by dividing the year up into its seasons and creates different prediction
depending on the dates of the data [29].
Astolfi et al uses several methods and comparisons including signal trending, physical models and ANFIS to compare two wind farms and with the results create models and KPI to assess the wind farms operations. The methods presented are both an indicator on what to expect from a wind farm before its being built as their KPIs can be used to assess different locations and also a tool for indicating when a wind farm is operating sub-par. The methods are made for assessment and approximation rather than an indicator of specific fault as most reports in the area are [19].
2.5. Literature review conclusion
Several methods have been tested when evaluating SCADA data for wind turbines. The trend shown in the reports indicates that the combination of several methods shows the most promise when finding errors or predicting wind turbine behaviour. The amount of data to analyse also creates new possibilities where machine learning and neural networks comes to the fore. Especially ANFIS is popular due to its computing speed, the ability to analyse new and unrecognised data and the possibility to help the program getting faster and more precise with logic rules based on experience. Another clear trend is that most research is done to predict faults in the components of the wind turbines. This leaves an area quite unresearched which is the finding of anomalies where the components are working and the alarm systems are silent. Some of the methods used in the presented reports could be used to find anomalies where the wind turbine underperforms without clear error. These could for example be due to bugs in the software or unresearched weather phenomenas.
3. OX2 and Wind Power
3.1. Wind turbines
There is a myriad of different designs of wind turbines as they have been developed over hundreds of years, but this thesis only handles the standard one. The three-bladed standard power producing turbine with a rated output of a couple of MW depending on exact size. For clearer reference in the report some of the parts of the wind turbine will be defined here with the help of Figure 3.2.
Figure 4 - Standard wind turbine design
1. Tower. The turbine tower is the structure which upholds the rest of the turbine. It also contains cables and such for information and power
transmission. The height of the tower is referred to as “tower height”.
2. Hub. The hub is the central part of the turbine which contains the generator, the gears, power transformation and most of the other technical systems of the turbine. The hub can turn on the tower and thereby have the ability to always face the wind. The hub direction is referred to as
“nacelle direction” and is often compared to the
“wind direction” which is also measured like an angle. Most of the meteorological measurement equipment are placed on the back of the hub, often in at least two sets.
3. Blade. The blades are the base component for energy transformation, from kinetic linear wind energy to kinetic circular mechanical energy that can be transformed into electrical energy by the generator in the hub. The base of these blades can also contain mechanical rotational possibilities to change the angle of the blade to be able to raise or lower the rotations of the blades. This system could also be used to brake the turbine in case of damaging wind speeds or technical failures. This blade angle is referred to as “pitch”.
3.2. The chosen wind turbine field
The method and the program developed will be tested on a single wind turbine farm that is operated by OX2. The owner of the wind farm and thereby the data has opted for anonymity so the wind farm, its location and the turbine data will not be presented in the report and the results will be adjusted for anonymity if required. For example, wind speed may be presented as a percentage of cut off wind speed for the turbines and so on. This will be presented in each section where anonymity is the case. The data from
the wind turbine field is from the year 2016 as this is the closest full year available.
3.3. Wind turbine type
The five chosen wind turbines are all the same model, which cannot be revealed due to company anonymity policy. It is a common type of wind turbine that is in use throughout northern Europe, on land and in medium wind conditions. The model includes an asynchronous doubly fed generator, a rotor diameter of over 100m, three blades, a gearbox with three gears, pitch controlled power and alternative current output. These specific turbines also include a modern anti-icing system that can heat the wings also when the turbines are operating to counteract icing on the blades. The model is operating in wind speeds between 2.5 m/s and 25 m/s with a rated output wind speed of 12.5 m/s.
3.4. Variables
The variables saved in the SCADA database are numerous but most of them aren’t relevant for this kind of study.
Some are measured as for example the component temperatures and others are calculated by the SCADA software to be used as indexes or to be easily used by operators or other software. The variable name could change between turbines as there are no standard between the manufacturers but the measured value is of the same type. Below are the main variables used presented with explanations of their origin and possible use.
3.4.1. Wind speed
As stated in the literature study, the wind speed has the greatest effect on the actual output of the turbine and even though its not the only changing variable that affects the output, its the major one. The wind speed is measured at the top of the turbine and are thereby affected greatly by the surrounding turbulence created by the wind turbine.
The error is predicted and the measurement adjusted to get a more realistic value but this adjustment is done where the turbine is produced and not at each actual location.
This makes the wind speed quite unreliable but close enough to use as a guideline for what the actual wind speed is. [35]
3.4.2. Ambient temperature
The ambient temperature around the wind turbine can affect the density of the air and thereby not only the kinetic energy but also the turbulence around the turbine.
The temperature is measured at the top of the wind turbine hub to get the most relevant one as it can differ from ground level temperatures.
3.4.3. Wind direction
The wind direction is measured at the hub together with the wind speed. This is a relevant variable as different wind direction can mean different turbulences due to surrounding terrain. The wind direction often changes from minute to minute, especially in turbulent condition and is just like the wind speed hard to measure correctly at the hub behind the blades.
3.4.4. Nacelle direction
The nacelle direction changes slowly to try to be straight towards the wind even though this is problematic due to the turbulent changing of direction. This variable is relevant due to the risk of not being synchronised with the wind direction but as long as they are somewhat aligned there shouldn’t be any problems.
3.4.5. Air pressure
Some turbines have pressure sensors at the hub to estimate the pressure of the surrounding air and thereby be able to calculate the density of the air going through the turbine.
Mostly does not affect the results to much but as its often correlated with the wind speeds and temperatures different variables can change together and thereby affecting larger changes in the output.
3.4.6. Active power
The active power of the wind turbine is the main variable targeted to measure the exactness of the program and the efficiency of the turbine. Its measured with high accuracy and the thing sold by the company and thereby the thing most important to maximize to get revenue. This will be the main tool used to calculate efficiency of the program and the error of the predictions.
3.4.7. RPM
The rotation per minute are a measure for how many rotations the wings are making each minute. Its measured exactly and closely related to the output but with a step function as the wind turbine contains gears that can lower the RPM but increase the output or vice versa at certain intervals. Could be used as a complement to measure the efficiency of the turbine
3.4.8. Pitch angles
The pitch angles are the angle of the pitch, the angle of the blades on the wind turbine. The turbine blades adjusted depending on wind speed and operation state to minimize or maximize the output of the turbine or help the turbine shut down in to strong winds. This variable could also be used as a tool for efficiency calculations as its closely tied to the wind speed and the output. Some wind turbines have different measurement points for different blades which means three different variables on a standard wind
turbine and often means they can be fined tuned depending on the characteristics of the wings. Other turbines have only a single measuring point and a simple angle for all the pitches of the turbine.
3.4.9. Operational state
Operational state is a variable that could mean different things for different turbines but most includes:
- Stop - Standby - Starting - Running - Free - Brake
These each have a number and depending on which one’s active the variable then contains the active number for that time step. This is not a measured point but rather the software mode for the state of the turbine.
3.4.10. Data availability and IEC-availability Like the operational state the data availability and the IEC-availability are something taken from the software of the turbine and can be used to filter certain time steps away to refine the results, more on this in the methodology section. There’s not always available data sent from the turbine to the database due to connectivity problems and the turbine itself does not have the capability to store data when the network isn’t working. When this data is lacking, the value is zero for this variable and when its available, its set to one and can thereby easily be filtered.
3.5. Data
The data downloaded from the SCADA servers can be downloaded as a text file or a “.mat” MATLAB file through Bazefield which is OX2’s interface system for the SCADA database. But it can also be directly imported through APIs as a struct file. The direct API download is a necessity in an automated program as its easy to define exactly which variable that should be imported from which time period. This also has the added benefit of speed as no interface is required to work as an
intermediary. This method means that own datafiles can be custom made to fit the latest settings chosen for the program for maximum efficiency. The data includes several things including the chosen variables for the chosen time period and each values’ respective time step.
The data doesn’t include variable names and other curiosities as the API import is very slim but such values are included in the import request and can thereby be organized in the same program structure as handles the rest of the API request.
4. Methodology
4.1. Basic calculations
This section presents the basic mathematical concepts used in the thesis to filter the incoming data or analyse the results produced.
4.1.1. Errors
The error analysis of the program is a central part and one of the most essential for the result. Several different kinds of error calculations are used to estimate the quality of the programs’ predictions. These are the absolute mean error, the error for each time step, the absolute error for each time step and the mean squared error. The different kinds of error are both used as a means of analysing the quality of the prediction made by the program by for example comparing the actual output compared to the predicted output, but also to as the analysis itself. If the error in a time period is much higher than the error in the adjacent time periods, there is probably something wrong and this could be one of the faults the program is made to find. The program has to be able to assess if the error between the prediction and the reality is just noise due to inconsistent variables and measurements or an actual error that has to be addressed. The programs error analysis use of these different errors is explained below.
4.1.1.1. Error and absolute error
The error and absolute error of each time step is used to calculate the quality of the prediction and find where the largest errors are.
#>OKHQPAANNKNL T F U
Equation 1 - Absolute error.
Where x is the actual output and y is the predicted at time step i.
4.1.1.2. Mean absolute error
The mean absolute error (MAE) is used as a simple single value to estimate how the whole process works. The mean of the absolute error for every time step is used instead of the mean of the errors as these are both positive and negative in equal measure and can cancel out each other which is misguiding. This mean absolute value is a good indicator for how well the prediction behaves. The goal is not to create a program with the lowest mean error as this would mean the faults in the wind turbine also are estimated. But still it has to be low enough to be relevant as to high an error would mean that the prediction is too bad to even estimate the common variable interactions. A balance has to be found and will be included in the testing and fine tuning of the programs options. The equation for the mean absolute error is presented below.
/#’ L s
J ˝ T F U
Æ
Equation 2 - MAE
Where n is the number of time steps, x is the actual output and y is the predicted output for each individual time step.
MAE is also used in to estimate the mean error over hours and days as that is a better indicator of actual faults than individual time step errors as these fluctuate greatly just as the variables does. But when these fluctuations continue to be of relevant size for an hour or even a day they are probably a good indicator for an actual error.
4.1.1.3. Mean squared error
The mean squared error (MSE) is used as a complement to the MAE to measure the quality of the prediction even though the result isn’t as simply used. The equation is presented below.
/5’ L s
J ˝ : T F U ;6
Æ
Equation 3 - MSE
Where n is the number of time steps, x is the actual output and y is the predicted output for each individual time step.
4.1.2. Standard deviation
Standard deviation (STD) is used in statistics as a measure of the dispersion of values in a data set. A high deviation measure means that the deviation of the data set is high and vice versa. Its calculated with the equation presented below.
56& L ' s
J ˝ :T F T§;6
Æ
Equation 4 - STD
Where n is the number of values in the data set, x is the individual value and T§ is the sum of all the values. The separate data values deviation can also be measured in standard deviations, using STD as a unit. This measure is used to know how far from the mean of the data set a value is set in regards to the spread of all the values. A STD from the centre of the spread is still in the group of variables as it per definition is the “standard deviation”, but five STDs from the centre is probably an outlier. As this measure takes into account the variable size and their spread, it is a good way of finding outliers in a random data set and can thereby be utilized in this thesis to sort away outlier values. Exactly how the STD is used for filtering is presented below in section 4.3.
4.1.3. Clusterification
Clusterification are the method of dividing points into different bundles depending on similarity or closeness in a multidimensional space. Which points that are paired depend on the clusterification method, if the clusters should be equal in amount of points, the number of clusters and if the variables are dimensionless. How this method is implemented are further explained in 4.2.1.
clusterdata and in 4.3.4.4. Cluster filtering.
4.1.4. Mahalanobis distance
The mahalanobis distance (MD) is the distance from the cluster measured in standard deviations rather than ordinary distance from the centre of the cluster. This makes it possible to easily identify outliers that deviate from the other samples. The standard deviation
calculations make it possible to find these even if the cluster of samples itself isn’t linear or spot formed and is thereby fitting to these kinds of wind turbine calculation. It will be used to sort outliers from the data set before the program should find out the standard variable interaction to get a more refined result.
4.2. MATLAB
The program is completely coded in the mathematical programing tool MATLAB. This is due to its suitable function and foundation for these kinds of automated calculations. The counter arguments to MATLAB are that it requires some complementing programming around it to make it able to automatically handle connection to the SCADA database but its positive characteristics makes up for this. Especially as these would have been needed anyway.
Below are the used advanced MATLAB functions presented and explained. Later in section 4.3 will their order and implementation also be explained.
4.2.1. clusterdata
The function clusterdata creates a specified number of clusters of the data input depending on the closeness of the samples place in a multi-dimensional space. These data clusters could then be used to manipulate certain parts of the data for example the data points at low wind speeds or when the active output is at maximum as the data would probably be clustered around these kinds of greatly changing variables. Single outliers in these clusters can then be filtered out as the probably depend on a single variable moving the sample point away from the cluster.
4.2.2. corrcoef
The function corrcoef calculates the correlation
coefficients for the variables. The correlation coefficients can be relevant when checking the variables for too close
dependencies for example between active output and rotations per minute as such dependencies could destroy the anomaly search in the result as the prediction also predicts some the faults and thereby does not find these.
The more samples the better correlation result can be calculated.
4.2.3. NN training
MATLAB has several neural network functions that together trains a neural network depending on chosen number of neural layers, limits, precisions and base theory. Its more of a program in the program than a single function and can be changed to fit whatever is needed. The type of neural network training used in the program will include the base theory of Levenberg-Marquardt backpropagation which is a least square fitting method used for nonlinear problems in computer aided
mathematical analysis. The main function used in the training is “train” which tries to use the chosen theory to find a pattern to fit a function after. The input of the function is the input chosen in the selection options and the result are the target or targets chosen. The generated neural network is rewritten in an iterative process and tested with the function “perform” to see which of the iterations that generates the better results. The more data input, the better the network, thereby should is it an advantage to use several months or even a year of data to get a general program with high precision. Another thing that can lower the precision of the program is lack of consistency in the variables. For example, the air pressure can be decreasing at the same time as the active power is increasing. These kinds of inconsistencies can make it hard for the program to create a precise NN and its important to limit the inputted variables as more variables means more inconsistencies between them. The end result of the training function is saved as a generated neural network function that can be called by the main program.
This function is then used by inputting new inputs and the neural network are given an estimation of what the target ought to be. This is an estimation that can be checked against the actual target variable of the turbine to see if the neural network or the turbine are behaving strangely. The neural network trained are made for a single purpose with the chosen inputs and targets. If a change is to be made to any of these or the option of the NN, the whole neural network has to be retrained. But the trained network can be used between different wind turbines with varying result. The neural networks produce better result the closer to the data used for the training the new data are. This means that its possible to gain efficiency in the program by using the same training on several turbines but with varying degree of precision. For best result the NN should be trained and used individually. The comparison between the individual networks and the more average network will be presented in the result section.
4.2.4. mahal
The function mahal does a mahalanobis distance
calculation on a cluster and presents the standard deviation
result for each sample. Good for outlier identification, especially as its not reliable on the size of each variable value or the size of the cluster.
4.2.5. listdlg
The function listdlg is a simple function that makes it possible for the user to easily choose variables from a list presented by an inbuilt graphical interface.
4.2.6. Plots
Three standard plot functions are used to present different results throughout the programs’ run. These are plot who creates a simple plot of a chosen variable over all time steps. Subplots who works similar but creates several parallel plots beside each other. Scatter3 who creates a scatter diagram in three dimensions to easily and graphically see the dependencies of three different variables to each other.
4.2.7. datetime
The function datetime is used to calculate the date and time of each corresponding anomaly found by counting the time steps from the data start date that comes with the standard input data file. These are a simple and efficient way of finding the correct date and time without having to program it yourself.
4.2.8. tic, toc
The function tic and toc starts and stops a timer and can be used to measure the used time of each function and thereby find and optimize different parts of the program for increased efficiency and thereby decreased runtime.
4.3. Logical structure
The MATLAB program will be created in two different versions with different goals. The first mode is created as a test platform for the efficiency and testing of different kinds of methods that could be used to analyse the operation of a wind turbine and how to best find the different anomalies. Then after this was achieved to a great enough degree the program is streamlined and automated in a second version created to automatically test data without any input more than the data needed for analysing and the chosen run options. Below are the different steps used in both versions of the program. The logical structure will be very similar with the main difference being that the automatic program doesn’t need any input other than a simple excel file while the test program needs a GUI with all possible options and a simple way of showing and saving the results.
4.3.1. Function and save based program The program is based on a function and save file
methodology that focuses on using few lines of code in an optimized central program that instead calls on the
functions needed who are coded separately to the core program. This way the program is faster and more optimized when used to achieve a single goal, which it often is if used in automation. All results and variables are saved in separate small .mat files between the different functions which minimizes the variables in the MATLAB workspace and active memory and further increases the efficiency and also makes it possible to access and use results from earlier runs and data sets. This function and save file methodology creates an efficient and greatly modular program structure that easily can be changed to suit different kinds of input data and the search for special anomalies. The program is created in this way to make it able to adapt to new and better analyse methods which is required to make the base program work with the new analytic methods that will be developed in the future and thereby prolong its efficient lifetime.
4.3.2. Options
In the start of the main program are a set of options that dictate which of the functions and limits the program should be using when it runs its analysis. These options make it possible to achieve different results with the same base program by for example optimize the process by limiting the functions used or make the program find different kinds of anomalies by choosing how the filters should be used or which accuracy the fault detection uses.
The base thought is that the program should be able to be adapted to the users wishes by only changing these options in the start of the base program. These options also include the choice of input and target variables for the program to use in the calculations. Different variable choices
completely change the NN and the results thereof.
4.3.3. API
The wind turbine data can be imported by connecting to the SCADA database through API (application
programming interface), a method which can be automated in MATLAB. The program will choose variables and the relevant time step for the data that’s needed according to its given parameters and can then use a dynamic import API to get the data from the SCADA database through use of the web. The data comes in the form of a simple struct file which can be translated into the program for easy use.
4.3.4. Preparation of data
The program uses several methods to prepare the data for the other functions. These includes mostly reading, understanding and then sorting the input data file into several sub-files for different uses so that the rest of the program easily can call for and get different variables.
This is one of the most important steps as the input data can look very different depending on how its imported.
Below are some of the main preparation methods used for different kinds of data.
4.3.4.1. Time
The time based data include different time steps, time periods and period length. In the preparation all these are read, understood and filtered down to their base
components in a uniform way. The time components that are identified and the saved into a time .mat file which includes time steps, start and stop time. It also creates an individual time step number vector that could be used in the filters to still be able to calculate the date and time when several days have been filtered away. Another thing created are a similar individual time step vector who are not filtered and the filtered values could be compared against.
4.3.4.2. Sorting
Sorting includes mostly reading the different kinds of variables in the data file and understands which vector who should be used as a base line for the filters that correlates to that variable later in the process. This are done by reading the individual identity name of the data vector and then in that name finding the included variable name. The sorting also includes finding which variables who are actual measured variables, which ones are software indicators of operation and which are unknown.
If an unknown is found, which could happen if a new turbine or a new naming system is used, the program could ask the operator for clarification or it could just treat it as an unknown variable and use it in the calculations.
4.3.4.3. Filtering
All the wind turbine data used by the program goes through several filters to achieve a better process. Some of the filters are used to get the program to work at all by filtering away faulty variables, for example NaN (not a number) variables and whole variable sections that for some reason doesn’t contain data. This is for the functions in MATLAB to be able to run at all as empty or faulty variables can stop the process completely as the function can not handle the input. Other filters are filtering away data that makes the training more precise in regards to what the target should be by filtering away for example low wind speeds, time steps with no data availability (where the SCADA has interpolated due to lack of information) and when then output are at the lowest or highest as these points probably aren’t containing any faults. There are several layers of filters and not all of them are used all the time. Different filters with different levels of accuracy can be changed in the option section of the program to help search for different faults or to try to achieve higher accuracy for a certain data set. But the first
filters mentioned are always used as the program wouldn’t be able to run its course at all with for example included NaN values.
4.3.4.4. Cluster filtering
Cluster filtering are set separate to the preparation of the data as it is the most time-consuming function and its only needed when training a new NN. Therefore, its put in the training function and are only used when needed to make the program faster and more efficient by not running unnecessary parts. The cluster filtering is also made dimensionless so as to give all the variables filtered the same weight. This is due to the predominance of large variables to otherwise dominate the clustering as its based on the sample inter-distance in the multidimensional plane. The cluster filter is based on the Mahalanobis distance explained in earlier sections. The maximal allowed mahalanobis distance is chosen in the options and the values that deviate from their cluster centres more than this are filtered from the data. This limit must be balanced as to filter away enough variables for the training data to be of a better quality and thereby create a better neural network. But it also should be low enough to not filter the needed correct data as more good time steps unfiltered gives better results.
4.3.5. Neural Network training
The neural network (NN) training are starting with the choosing of input and target choices which the human operator can handle through the simple listdlg function.
The program is then using these choices along with some options chosen at the start to train the NN with the inbuilt MATLAB functions. The trained program is continuing this process until a limit is reached or a high enough degree of fitting has been achieved and is then saving the NN as a separate function that can be used by the program later for this input data or another run of the program with different data. The thought is to train the NN with a whole year of historical data and then use this NN to test days and weeks continuously with a fast already finished and trained NN. The selections in the start of the training is saved in a separate file so that the program automatically can choose the correct variables when used on another data file. This is important as the trained program only works with the same inputs and targets.
4.3.6. Error estimation
A prediction is created by the program for each of the targets chosen by analysing the input variables with the NN. The prediction is then compared to the actual targets also saved in the data file and the errors are found and saved in separate files as they are for each time step but also as mean and absolute values for further analysis.
Especially the absolute values are interesting as the
heavily fluctuating values often takes each other out when using mean values. The absolute values are then calculated as a mean absolute value for both days and hours. Hours are the most interesting value as this is a good estimation on the prediction quality for each hour at the same time as small errors for single time steps disappears when the mean is taken. In the options in the start of the program a mean standard deviation of the error is chosen as a limit of the errors and when the mean absolute error for an hour passes this limit the program saves this time step ID for further analysis. The standard deviation makes the program flexible in terms of different variables and also adapts to the quality of the prediction which makes the whole program more versatile. These found large errors are then presented to the user with a simple zoomed in graphical plot together with data about the error length in minutes, the date and time of the error and the error size in KWh. The user can then easily see if the error found is an actual interesting error or if its a false alarm. If there’s to many false alarms the program is meant to be tweaked with the options and filters to achieve a higher accuracy and skip over the false alarms of the data but this could differ between turbines and wind farms.
4.3.7. Training rerun
A large problem with the neural networks is that the network can not differentiate between good and bad data, this leads to training on time steps with large errors due to for example icing which leads to an overall worse neural network. A method that may help to solve this problem is a simple rerun of the training after an additional filter that removes all time steps with found errors from the first run.
This would mean that all potential errors in the first run is removed and thereby would the second trained network not include all the large errors in the data in its training and thereby theoretically be significantly more precise in its predictions of the errors. This could theoretically mean that smaller errors that was unnoticed in the first neural network is found by this second, better one. This will be tested and the results presented in the result section.
4.3.8. Automation
In the end, the goal of the program is automation of data analysis and thereby when the program has done all the calculations, it saves all the variables, results and relevant graphs into a save folder with the same name as the base data file. Thereafter it takes the next wind turbine data file on the list and begins anew from the top and continuous to do this until everything is tested or a user intervenes.
4.4. Setup of program testing
There are several possible ways to utilize the different methods and their settings in the program, therefore a
number of test runs will be done to find the optimal solution and design of the program. Five wind turbines have been chosen by OX2 for their normal, not perfect, behaviour. As the anomalies dominating the turbine operation in winter time are due to ice on the blades, the focus months chosen are the summer half of the year. This is due to other alarm system that is indicating icing which means the program would find more relevant anomalies in summertime. All data tested and used to train the neural network comes from 2016. The program uses the whole year to train on but only looks for anomalies in the summer, April to September.
4.5. Result analysis
The analysis of the results and the quality of the utilized method will be evaluated in two ways. Firstly, by the accuracy of the prediction in regards to the actual target values, as this is a good quantifiable quality result. But the accuracy isn’t enough as the goal is to find the anomalies in the wind turbines’ operation. As mentioned above, a perfect accuracy would mean that no anomalies were found. Therefore, the second method will complement the accuracy by using a confusion matrix to further analyse the result [36]. The confusion matrix consists of four categories displayed in figure 5:
Figure 5 - Confusion matrix
TP is the correct found fault in the data while TN are the rest of the data set not recognized as an anomaly by the program and not containing any actual faults. FP is the faults predicted which aren’t actual faults but just faulty predictions. FN are the anomalies that aren’t found by the program for different reasons and the ones that needs to be identified so that the program can be redesigned to include these also. This confusion matrix analysis will be made for each hour as the anomaly prediction works with hourly time steps. The percentage of correctly predicted anomalies will then be calculated by the following to equations:
