Issue link: https://industrialengineer.epubxp.com/i/920036
52 ISE Magazine | www.iise.org/ISEmagazine Inside IISE Journals research Understanding the wake effect requires physics-based and data-driven modeling The wake effect is a unique aerody- namic feature that happens when wind passes through a turbine. After flow- ing through the rotating wind turbine blades, the increased turbulence and decreased energy in the wind flow can curtail power generation by 50 percent, causing a wind farm to lose more than $1.3 million of revenue each year. To alleviate the power loss in wind turbine operations, the wake effect should first be characterized and quan- tified. Sophisticated computational fluid dynamic models help understand the wake effect, but their prohibitive com- putational demands mean they are not practical for use at commercial-size wind farms. The wake effect is typically esti- mated either through simplified physical models or based on purely data-driven methods. Both approaches are less than ideal. The physical model overlooks useful in- formation in real-time operational data while noisy measurements significantly affect the purely data-driven approach. In "Spline model for wake effect anal- ysis: Characteristics of single wake and its impacts on wind turbine power gen- eration," Hoon Hwangbo, a postdoc- toral researcher, along with professors Andrew Johnson and Yu Ding, all from Texas A&M; University, present a sign- constrained statistical learning method for quantifying the wake effect charac- teristics. It is a hybrid of physical knowl- edge and data science modeling; the learning based on spline models comes from data science while the sign con- straint is derived from physical under- standing. In an out-of-sample test con- ducted on six pairs of wind turbines, the proposed hybrid approach reduced pre- dictive error by 30 percent over a widely used physical model and by 6 percent over the purely data-driven model. This 6 percent reduction in predictive error is strong evidence of overfitting commit- ted by the purely data-driven approach. Overfitting refers to the circumstances This month we highlight two articles from IISE Transactions. The first article develops a sign-constrained statistical learning method by innovatively integrating physical knowledge and data science modeling methods for quantifying the wake effect characteristics. The proposed method is used to predict the wake effect of wind turbines with real data sets and demonstrated significant improvement in reducing prediction error compared with only using physical model or purely data-driven methods. The second article develops a connected-path filtering method to detect systematic patterns of wafers, which helps find the root causes of failure to find the right interventions for quality management. The method was tested on real wafer bin map data from SK hynix with great success. These articles will appear in the February 2018 issue of IISE Transactions (Vol 50, No. 2). Yu Ding (from left), Hoon Hwangbo and Andrew Johnson show off a computer screen that has a visualization of their wake effect quantification.