Enhanced Turbine Performance Monitoring 08-11_Iowa

7/29/2019 Enhanced Turbine Performance Monitoring 08-11_Iowa

1/6

36 AUGUST | 2011

EnhancEd TurbinE

PErformancE moniToring

comPonEnTs of wind TurbinEs are aectedby asymmetric loads, variable wind speeds, and se-

vere weather conditions which cause wind turbinesto change their states. A typical wind turbine under-goes various states during its daily operations. Tewind turbine states ollow a certain pattern, such as:

1) turbine OK to run-up idling; 2) turbine online tomaintenance/repair mode; 3) turbine weather con-ditions to external stop, and; 4) turbine OK to aultmode, and so on. A state change rom normal tur-bine operations to a ault mode adversely impacts theperormance o wind turbines and its components.

Monitoring these states can greatly augment themaintenance operations.

Wind turbine monitoring can be done in two ways:condition-based, and perormance-based. Condi-tion monitoring requires installation o additionalequipment/sensors to continuously monitor relevant

parameters in real time. Perormance monitoring uti-lizes historical data or prediction o turbine peror-mance. Since perormance monitoring relies on theexisting data, there is no additional cost to the windarm operators. Analyzing the historical data throughdata-mining techniques is a promising approach to

The authors o this article clearly demonstrate the

benefts o applying data-mining techniques to theprediction o ault modes in wind turbines.

By Andrew Kusiak and Anopp Verma

ae Kk ty app Ve te tet te deptet me it E-ee t Te uvety i. Kk e ee t [email protected]. a vt ..e.


2/6

windsystemsmag.com 37

perormance monitoring. Te current SCADA sys-tems already record wind turbine parameters. Animproved SCADA system coupled with data-miningalgorithms can be useul in identiying critical per-ormance indicators o wind turbines. Tis articledemonstrates application o data-mining techniques

to prediction o ault modes o a wind turbine.

TurbinE sTaTEs informaTionIn addition to wind turbine parameters, a SCADA systemrecords the states o wind turbines. Consider 17 possiblestates in which a turbine can be ound. Te state tur-

bine in ault mode (state No. 17 in able 1) can usuallybe expressed in more than 400 ways. Components oa wind turbine such as wind turbine blades (e.g., bladeangle asymmetry), turbine yaw (e.g., yaw runaway),wind turbine generator (e.g., generator brush worn),and wind turbine gearbox (e.g., gearbox over-temper-ature) can be aected. Using domain knowledge, thestates listed in able 1 can be broadly categorized intoour main groups (g. 1). At the top level, prediction o

ault modes o wind turbine is crucial in identiying anactual ault.

Figure 2 displays the distribution o wind turbinesstates. Te values are averaged or a wind arm over a

year period. Almost 90 percent o the time the turbine isoperating normally. However, aults make up 6.82 per-

f. 1:ctee tte te.

f. 2:dtt te tte.


3/6

38 AUGUST | 2011

cent o the total time. Any eort to minimize the requency o aultswould enhance wind turbine availability.

dEVEloPing PrEdicTion modElsIn this section, data-mining models are developed or predicting tur-bine states. he data collect by the SCADA system is sampled at therequency o 0.1 Hz. he aults that have occurred at 17 wind turbinesin a three-month period are plotted in ig. 3. Based on the requency

o ault modes, three wind turbines are identiied (labeled as A, B, andC in ig. 3). Data rom turbine A is used or training and testing data-mining algorithms, whereas turbine B and C are used or perormanceevaluation.

ParamETErs forsTaTEs PrEdicTionA SCADA system records vari-ous parameters, which can becategorized into: 1) non-con-trollable parameters such aswind speed and wind devia-tions; 2) perormance param-

eters such as power output androtor speed; 3) vibration pa-rameters such as tower accelera-tion and drive train acceleration,

State

Number

State Description State

Number

State Description

1 Turbine OK with no errors 10 Turbine stopped locally

2 Turbine running smoothly 11 Emergency stop

3 Turbine running up idling or cut in 12 Turbine stopped due to curtailment

4 Turbine in maintenance mode 13 Turbine stopped by customer

5 Turbine in repair mode 14 Turbine idling locally

6 Power ailure/grid downtime 15 Turbine idling remotely

7 Weather downtime 16 Wind direction curtailment

8 Turbine stopped externally 17* Turbine in ault mode

9 Turbine stopped locally

* Primary ocus

Te 1: Te tte t.

awEa s &ctywpe

Pee vt tbt #127


4/6


Te evaluation o data-mining algorithms is based on the prediction accuracyo turbine states namely turbine OK (O), weather downtime (WD), mainte-nance downtime (MD), and ault mode (FM) (g. 4). Te diagonal elements (e.g.

and; 4) temperature parameterssuch as gearbox temperature andgenerator temperature. It is im-portant to mention that not allthe parameters are causing wind

turbines to change its states.Tereore, we need to identiy aset o parameters that impact thewind turbine states. Tis is theplace where algorithms build us-ing data-mining techniques areuseul. Data-mining algorithmsuse statistical measure such asinormation gain and correlationcoefcient to identiy relevantset o parameters. able 2 shows

10 parameters relevant to windturbines states (data rom tur-bine A), identied by three di-erent data-mining algorithms.Te techniques used are wrap-per with genetic search (WGS),wrapper with best rst search(WBFS), and boosting tree al-gorithms (BA). For detaileddescription o these techniquesreer to [1-3].

daTa-mining modElsA combination o relevant inputparameters (ound in previoussection) can be used to developprediction models or state aults.

No. WGS WBFS BTA

1 Nacelle revolution Blade 3 pitch angle(actual) Blade2 pitch angle (actual)

2 Blade 3 pitch angle (actual) Current phase C Blade3 pitch angle (actual)

3 Current Phase B Temperature hub Blade1 pitch angle (actual)

4 Nacelle Position Temp. control box axis 1 Generator/gearbox speed

5 Generator/gearbox speed Voltage phase C Generator speed

6 Temperature, bearing B Generator speed Rotor speed

7 Temperature top box (C) Drive train acceleration Blade2 pitch angle (set)

8 Power (Actual) Temperature top box Blade3 pitch angle (set)

9 Tower deection Nacelle revolution Blade1pitch angle (set)

10 Wind deviation, 1 sec Temperature bearing A Drive train acceleration

Timestamp [s]

Turbine state OverallAccuracy [%]Turbine OK

[%]Fault mode

[%]Maintenance

downtime [%]Weather

downtime [%]

t 99.88 99.67 78.41 97.91 99.45

t + 10 99.56 99.00 77.22 95.04 98.39

t + 30 97.64 96.41 74.59 94.62 96.54

t + 60 95.70 95.64 71.67 92.55 94.43t + 120 91.87 90.00 67.49 88.47 90.89

t + 180 88.58 87.34 64.94 84.43 86.82

t + 240 85.62 84.64 60.31 82.44 83.93

t + 300 83.05 82.76 59.67 80.39 81.76

Actual state Anticipated state Correctly identifed cases

Blade angle not plausible axis 2 Fault 76.92%

Gearbox oil pressure too low Fault 0.00%

Maintenance downtime Maintenance downtime 100%

Motor protection Fault 100%

No activity CAN-Bus CCU Fault 50.0%

Overproduction Fault 100%

Pitch control deviation axis 3 Fault 100%

Saety chain Fault 100%

Turbine OK Turbine OK 99.45%

Weather downtime Weather downtime 60.70%

Yaw runaway Fault 87.50%

Te 2: Pete eete t t- t.

Te 3:Pet

y te

tte(te a).

Te 4:ay te b.

f. 3: feqey t e te.


5/6

40 AUGUST | 2011

PO, PWD) are correctly predicted turbine states,whereas non-diagonal elements are wrongly predicted

(e.g., FPFM, MD: turbine in ault mode is predicted asturbine in weather downtime).

Data-mining algorithms namely random orest algo-rithm (RFA) [4] is used to build eight prediction modelsat various time stamps, with the maximum predictionlength o 5 min. Te selection o data-mining algorithmis based upon their perormance on training data attime stamp t. Te accuracy was ound to be in the range81-99 percent or all turbine states (able 3).

robusTnEss of daTa-mining modElsIn order to check the robustness o developed models,

various unseen aults were tested. Te aim is to check the

response o the model when unobserved aults are en-countered. Figure 5 (a-b) displays the distribution statesor turbines B and C plotted over a three-month period.Te number o aults varies across turbines, however, tur-bines are ound to be operational most o the time. Teresults shown in ables 4 and 5 illustrate the response odata-mining algorithms on turbine A and B. Except o thegearbox-related aults, other ault modes o wind turbinesare correctly identied.

Actual state Anticipated state Correctlyidentifed cases

Centriugal switch Fault 100%

Gearbox oil over-temperature Fault 0.00%

Maintenance downtime Maintenance downtime 100%

Pitch overrun 00 Fault 100%

Power ailure Weather downtime 60.00%

Turbine OK Turbine OK 99.27%

Weather downtime Weather downtime 95.54%

Yaw runaway Fault 100%

f. 4:ay eet te tte.

ModularDrawerSystem

STOR-LOC

MODULAR

DRAWER

SYSTEM

880 N. Washington Ave.

Kankakee, IL 60901

Toll Free: 1.800.786.7562

Fax: 1.800.315.8769

email: [email protected]

www.storloc.com

Te 5:ay te c.


6/6


conclusionTe parameters recorded by SCADAsystems can be useul or monitoringpurposes. Using prediction models

derived by data-mining algorithmsthe states o wind turbines can be pre-dicted ahead o the time, which canbe helpul in maintenance planning.Te models built using data-miningalgorithms can be integrated withcurrent SCADA system to enhanceperormance monitoring o windturbines. Acknowledgement: Te re-search reported in the paper has beensupported by unding rom the Iowa

Energy Center, Grant 07-01.

rEfErEncEs:1) Kohavi, R. and John, G.H., 1997,

Wrappers or eature subset se-lection, Articial Intelligence,97(1-2), pp. 273-324.

2) Sbihi, A., 2007, A best rstsearch exact algorithm or the

Multiple-choice Multidimen-sional Knapsack Problem, Jour-nal o Combinatorial Optimiza-tion, 13, pp. 337-351.

3) Kudo, . and Matsumoto, Y.,2004, A Boosting Algorithm or

Classication o Semi-Structuredext, EMNLP.

4) Liwa, A. and Wiener, M., 2002,Classication and regression byrandom orest, R News, 2(3), pp.18-22.

Learn more about SEL Engineering Services

at www.selinc.com/wind.

Managing wind energy requires innovative solutions. From electricalengineering studies and design to installation and testing, the SELEngineering Services team provides superior quality and performanceusing the best products in the industry.

Complete system protection and data integration

Automated megawatt output load control to interconnecting utility

Revenue metering with the ability to serve multiple owners

Automated capacitor control

Get Connected. Stay Connected.SEL Engineering Services and Consulting

[email protected] | +1.509.332.1890

f. 5: stte tt te: () Te b, () Te c.

Documents

Enhanced Turbine Performance Monitoring 08-11_Iowa