EXPLORING KEY VARIABLES IN WIND TURBINE POWER CURVE MODELING A Thesis by DAVID MATTHEW PEREZ Submitted to the Office of Graduate and Professional Studies of Texas A&M University in partial fulfillment of the requirements for the degree of MASTER OF SCIENCE Chair of Committee, Yu DingCommittee Members, Li Zeng Mladen Kezunovic Head of Department, Mark Lawley August 2018 Major Subject: Industrial Engineering Copyright 2018 David Matthew Perez
ii ABSTRACT Though substantial evidence has shown the importance of wind speed and direction in modelling a wind turbine’s power curve, there remains uncertainty as to whether other variables would improve modelling efforts and which specific variables those would be. This present work expands upon prior research using the additive multiplicative kernel (AMK) technique to explore the use of additional variables. The experimental methodology involves arriving at power estimates by treating a year’s amount of wind turbine data as a learning problem. Different combinations of variables are investigated and compared in terms of error reduction on the testing set using root mean square error and mean average error. Discussion on the best sets of variables combinations are presented to gain insight as to why certain variables lead to greater error reduction and whether they are likely to be included in different sets of data. Two categories of variables emerge from the research. The first includes variables that are typically recorded in field operations including time, turbulent intensity, and the standard deviation of wind direction. Time and turbulent intensity are shown to offer promising results. The next set includes variables that measure how much wind speed and direction vary as height varies. This is especially of interest as turbine size has increased substantially over recent years. In particular, the rotor equivalent wind speed neatly captures the variation of wind speed and direction across the length of a turbine’s rotor in a single value. Using this parameter with AMK leads to significant prediction error reduction, making a strong case to include it in modeling the power curve. As will be discussed, doing so proves to be a better alternative than current industry practice.
iii ACKNOWLEDGEMENTS I would like to thank my committee chair, Dr. Ding, and my committee members, Dr. Zeng and Dr. Kezunovic for their guidance in making this work of research possible. Dr. Ding’s support enabled me to fully focus on classes and research. Support also thankfully came from the Industrial & Systems Engineering Department in my time as a teaching assistant. I owe a tremendous debt of gratitude to Texas A&M University for their generous financial support in the form of the Graduate Diversity Fellowship. Without the fellowship, it is doubtful if I would have been able to pursue a Master’s Degree, so I will be eternally grateful to the University for giving me this opportunity of a lifetime.