Wind power provides 20 percent of renewable energy in the United States. Turbines to harvest the wind usually are built in rural or open-water areas to use the wind efficiently and generate the greatest amount of energy; however, reaching and repairing individual wind turbines is time-consuming and expensive. A small cluster of malfunctioning wind turbines may significantly reduce crucial energy output.
In extreme cases, the entire wind turbine may fall off the wind-power generator, causing damage to other expensive equipment and its surroundings. So, there is a great sense of urgency in the wind energy industry to develop pre- dictive models to forecast which components in which wind turbine may fail, or even which repair parts are needed.
Because my research focuses on data mining and machine learning, I was invited by a wind energy consulting firm to develop machine-learning models to predict the failure of wind turbine components. My research team consists of three international students from the Graduate Programs in Software: Kiran Guntupalli from India, Christian Klaue from Germany and Ruogu Wang from China.
Machine-learning methods build predictive models by discovering associations between wind turbine sensor history and wind turbine failure history. Because each wind turbine can have more than 100 sensors attached to it, and each sensor may collect data every 1 to 10 seconds continuously over a very long time period, the data that needs to be processed and “learned” by machines is enormous. In fact, this project has more than 7 terabytes of data that were collected over a three-year period from hundreds of different types of wind turbines.
To speed up these complex machine-learning processes, my team and I plan to use the Amazon cloud computing platform and graphic processing units. Currently, the predictive models we’ve built have achieved about 80 percent accuracy in predicting wind turbine failures.
However, because the operation environment of wind turbines may keep changing (i.e., increasing temperature due to global warming, or gradual deterioration of components), the team is planning to extend its current models so they can update themselves dynamically in real time based on drifting data.
