Data Analytics for predictive maintenance of Wind Turbines.
| dc.contributor.advisor | Basson, A. H. | en_ZA |
| dc.contributor.advisor | Kruger, K. | en_ZA |
| dc.contributor.advisor | Grobler, J. | en_ZA |
| dc.contributor.author | Devenish, A. | en_ZA |
| dc.contributor.other | Stellenbosch University. Faculty of Engineering. Dept. of Mechanical and Mechatronic Engineering. | en_ZA |
| dc.date.accessioned | 2021-11-30T20:56:26Z | |
| dc.date.accessioned | 2021-12-22T14:26:08Z | |
| dc.date.available | 2021-11-30T20:56:26Z | |
| dc.date.available | 2021-12-22T14:26:08Z | |
| dc.date.issued | 2021-12 | |
| dc.description | Thesis (MEng)--Stellenbosch University, 2021. | en_ZA |
| dc.description.abstract | ENGLISH ABSTRACT: There is a global drive for using greener energy sources for power generation, for which wind energy is a popular alternative. Significant investment is thus attracted to wind farm development. However, research into wind farm management is required to improve the economic viability of wind turbines, and maintain wind energy as a competitive source for electricity generation. Elements of wind farm management that contribute significantly to the overall costs of operating wind farms, are maintenance and repair operations. Maintenance and repair operations constitute 20-25% of the total levelised cost of wind turbines. Decreasing these costs would thus greatly contribute to improving the operating costs of wind farms. The research in this thesis investigates the detection of wind turbine failures to aid the contribution of preventative measures that can be taken to decrease the total levelised cost of wind turbines. The objective of this thesis is to evaluate the use of data analytics for predictive maintenance of wind turbines. Therefore, the thesis presents a review of predictive maintenance solutions found in literature, as well as a case study which demonstrates the use of various data analytic techniques for predictive maintenance of wind turbines. In the case study, various machine learning algorithms are evaluated for predicting failures in wind turbines by two broad approaches - classification and regression. The classification approaches predict whether a failure will occur in a certain time period, while regression approaches estimate the time until a failure occurs. The best performing algorithm for classification in the case study was the support vector machine (SVM), and for regression the random forest. The classification and regression models were evaluated using recall and precision, and root mean squared error (RMSE), respectively. Binary and multiclass experiments were performed for both classification and regression approaches. The binary predictions provided failure warnings without an indication of which component was going to fail, whereas the multiclass experiments made separate failure predictions for each component category. A conclusion drawn from the multiclass predictions is that the failure categories have different predictabilities, as some showed to be more successfully predicted than others. Knowing which component is going to fail is valuable to wind farm operators because they can immediately focus on implementing preventative measures for that specific problem in the turbine. The time required to locate the cause of the failure warning, inspect and fix what is faulty, and return the turbine to normal operation again, will therefore be reduced. From the outcomes of this thesis, it is concluded that for the dataset used in this research, regression is the most reliable approach for predicting failures in wind turbines. Even though the multiclass cases' data was more severely imbalanced than the binary cases, the multiclass failure predictions were better for some components than the binary case predictions: the binary case's RMSE was 30.9 hr, while the multiclass errors ranged from 10.2 hr for transformer failures to 21.4 hr for hydraulic group failures. | en_ZA |
| dc.description.abstract | AFRIKAANSE OPSOMMING: Daar is 'n wêreldwye strewe na die gebruik van groener energiebronne vir kragopwekking, waarvoor windenergie 'n gewilde keuse is. Aansienlike belegging word dus aangetrek deur die ontwikkeling van windplase. Navorsing oor bestuur van windplase is egter nodig om die ekonomiese lewensvatbaarheid van windturbines te verbeter en windenergie as 'n mededingende bron vir die opwekking van elektrisiteit te handhaaf. Elemente van windplaasbestuur wat aansienlik bydra tot die algehele koste van die bestuur van windplase, is onderhouds- en herstelwerk. Onderhouds- en herstelwerk vorm 20-25% van die totale gelykmatige koste van windturbines. Die verlaging van hierdie koste sal dus 'n groot bydrae lewer tot die verbetering van die bedryfskoste van windplase. Die navorsing in hierdie tesis ondersoek die opsporing van windturbine-falings om die bydrae van voorkomende maatreëls wat geneem kan word om die totale gelykmatige koste van windturbines te verlaag, te help. Die doel van hierdie tesis is om die gebruik van data-analise vir voorspellende instandhouding van windturbines te evalueer. Daarom bied die tesis 'n oorsig van voorspellende instandhoudingsoplossings wat in literatuur voorkom, asook 'n gevallestudie wat die gebruik van verskillende data-analitiese tegnieke vir voorspellende instandhouding van windturbines demonstreer. In die gevallestudie word verskillende masjienleer-algoritmes geëvalueer om falings in windturbines te voorspel deur twee breë benaderings - klassifikasie en regressie. Die klassifikasiebenaderings voorspel of 'n faling in 'n sekere tydperk sal plaasvind, terwyl regressiebenaderings die tyd skat totdat 'n faling plaasvind. Die algoritme wat die beste gevaar het vir klassifikasie in die gevallestudie, was die "support vector machine" (SVM) en vir regressie die "random forest". Die klassifikasie- en regressiemodelle is onderskeidelik geëvalueer met behulp van herroep ("recall") en wortel-gemiddeld-kwadraat-fout (WGKF, "root mean squared error", RMSE). Binêre en multiklas-eksperimente is uitgevoer vir beide klassifikasie- en regressiebenaderings. Die binêre voorspellings verskaf falingswaarskuwings sonder 'n aanduiding van watter komponent gaan misluk, terwyl die multiklas- eksperimente afsonderlike falingsvoorspellings vir elke komponentkategorie gee. 'n Gevolgtrekking wat gemaak word uit die voorspellings van meervoudige klasse is dat die falingskategorieë verskillende voorspelbaarhede het, aangesien sommige meer suksesvol voorspel kon word as ander. Om te weet watter komponent gaan misluk is waardevol vir windplaasoperateurs omdat hulle onmiddellik kan fokus op die implementering van voorkomende maatreëls vir die spesifieke probleem in die turbine. Die tyd wat nodig is om die oorsaak van die foutwaarskuwing op te spoor, te inspekteer en reg te stel wat foutief is, en die turbine weer in normale werking te bring, sal dus verminder word. Uit die uitkomste van hierdie tesis word die gevolgtrekking gemaak dat regressie vir die datastel wat in hierdie navorsing gebruik is, die betroubaarste benadering is om falings in windturbines te voorspel. Alhoewel die gegewens van die multiklasgevalle ernstiger in wanbalans was as die binêre gevalle, was die voorspellings in multiklasgevalle beter vir sommige komponente as die voorspellings in die binêre geval: die WGKF van die binêre geval was 30.9 uur, terwyl die veelklasfoute gewissel het van 10.2 uur vir transformatorfoute tot 21.4 uur vir foute in die hidrouliese groep. | af_ZA |
| dc.description.version | Masters | en_ZA |
| dc.format.extent | 140 pages | en_ZA |
| dc.identifier.uri | http://hdl.handle.net/10019.1/123869 | |
| dc.language.iso | en_ZA | en_ZA |
| dc.publisher | Stellenbosch : Stellenbosch University | en_ZA |
| dc.rights.holder | Stellenbosch University | en_ZA |
| dc.subject | Predictive maintenance | en_ZA |
| dc.subject | Failure analysis (Engineering) | en_ZA |
| dc.subject | Data analysis | en_ZA |
| dc.subject | Wind turbines | en_ZA |
| dc.subject | UCTD | en_ZA |
| dc.title | Data Analytics for predictive maintenance of Wind Turbines. | en_ZA |
| dc.type | Thesis | en_ZA |