skip to primary navigationskip to content

MPhil in Engineering for Sustainable Development

global challenges, engineering solutions

Studying at Cambridge

 

Kaustav Dutta

Accurate prediction of wind turbine component failure using predictive data analytics: Investigating methods for determining wind turbine health

Existing research for predicting wind turbine component failures (WTCFs) has primarily concentrated on Supervisory Control and Data Acquisition (SCADA) signals relating to power output, machine vibrations and component temperatures. The performance of a wind turbine (WT) is largely governed by its behaviour in relation to the incident wind. This behaviour is revealed by the analysis of pitch and yaw systems signals. Methods for predicting WTCFs based on the analysis of pitch and yaw systems, however, are limited. Recent evidence suggests that these methods that make use of leaner datasets are both computationally and economically cost effective compared to traditional methods.

This dissertation outlines the design of a multi-feature integrated system that is capable of predicting future pitch and yaw values and classify turbines into faulty and healthy states. These predictions are based on historical learnings and deviations from normal behaviour.

Firstly, an in-depth literature review has been undertaken in order to; identify prevalent techniques of data management, data selection and data pre-processing, and compare the performance of several pattern recognition techniques.

Following this, a robust methodology has been developed utilising an epidemiological approach to determine the health of WTs. This has added to the novelty of the work carried out here, as such an interdisciplinary approach is rare in this field. The results of the case-control study carried out show that faulty turbines exhibit significantly high variance in pitch and yaw movement, compared to healthy turbines.

Finally, the regression and classification models have been designed and evaluated individually before being integrated into a final model. The results of these evaluations show that Long Short-Term Memory (LSTM) networks coupled with a Random Forest classifier are capable of predicting future pitch and yaw signals of a turbine and determining their operational state; faulty or healthy. This model exhibited fault visibility up to 3 weeks before the manifestation of the fault on the WTs tested.

The methods developed here should be validated further by introducing more failure related data and incorporating selected SCADA signals to reveal the root cause of failures. In particular, temperature-based analyses coupled with effective SCADA alarms handling should be considered for increased granularity in fault predictions.