Novel Function Estimation and Comparison Methods with Applications in Wind Power Curve Modeling

Abstract

Wind energy is the forerunner among the renewable energy sources, and by the end of 2020, wind energy accounted for roughly 8.4% of the total electricity used in the United States. In various decision making tasks related to wind farm operations, including wind power prediction and turbine performance evaluation, wind power curve plays an important role. A power curve is a function that maps the relationship of wind speed and other environmental variables to the wind power output. This work focuses on developing new data-driven methodologies to better model and compare power curves. The dissertation tackles three distinct problems related to power curve modeling.

The first problem is related to temporal autocorrelation in wind turbine data. The autocorrelation results in a phenomenon called temporal overfitting, which has gone unnoticed in power curve modeling literature. This work highlights the temporal overfitting problem in wind power curve modeling and proposes a remedy for the same by splitting the power curve into a time-independent and a time-varying component. The time independent component is then used for temporal extrapolation for a new time period.

The second problem is about comparing two power curves given their noisy datasets. Till now, power curves were only compared using point metrics. This dissertation proposes a functional hypothesis test to compare the power curves throughout the input space and get a more holistic comparison. A key feature of our proposed method is identification of input regions where the functions under comparison are different.

The last part of the dissertation jointly models the power curves of an entire wind farm as a function of the underlying terrain characteristics. This modeling approach provides insights into the effect of terrain on power production. It also improves the transferability of power curves to a new site for which one only has the terrain information.

Overall, the methodologies adopted in this dissertation yield significant improvement and push the boundary of knowledge for different aspects of power curve modeling.