Integrating yaw misalignment in power curve modeling for wind turbines: a data science approach

The rise of global warming and need for sustainability have placed an emphasis on clean, inexhaustible, and renewable sources of electric power. Wind energy has proven to be promising in this regard, being one of the largest renewable sources of electricity in the U.S. and worldwide, owing to technological advancements and its continuously falling costs. For wind energy, the fundamental relationship governing the wind-to-power conversion is called a “power curve,” which relates the power output of a wind turbine with the wind conditions unfolding at the turbine’s hub height. A power curve enables the estimation and prediction of power produced by a wind turbine, along with serving other important purposes, ranging from performance monitoring and productivity assessment to maintenance scheduling and optimization. The International Electrotechnical Commission (IEC)’s “method of bins” is the current industrial standard for power curve modeling. While prevalent, the method of bins suffers from a number of limitations, among which, is its dependence on the wind speed as the sole input. Therefore, an active research area in recent years has sought data-driven methods that, in addition to wind speed, can include the impact of other secondary, yet potentially significant factors, such as turbulence intensity, air density, humidity, air pressure, among others. Nevertheless, little research has been performed on the incorporation of operational variables, such as yaw misalignment, into power curve models. Yaw misalignment is the relative direction of the incoming wind with respect to the axis of the wind turbine and is known to significantly affect the power output of a wind turbine. In this study, the potential of incorporating yaw misalignment into power curve modeling is studied. In specific, a novel data-driven approach, called the “Smart Neighborhood Search,” or in short (SNS), has been proposed for power curve modeling. Instead of directly accounting for yaw misalignment as an additional input in a bivariate power curve formulation which has been shown herein to offer little to no competitive advantage, SNS constructs a set of local conditional polynomial models wherein the relationship between yaw misalignment and power output is locally learnt, conditional on a given wind speed. Two variants of this model have been proposed, and their performance is compared with other power curve modeling approaches, including the method of bins, along with other prevalent machine learning approaches such as k nearest neighbors (kNN), random forest, additive multivariate kernel, and neural network. Tested on a real-world dataset from two wind turbines located at Engie’s La Haute Borne wind farm, the SNS approach is found to outperform existing power curve models, in terms of several prediction accuracy metrics. This study opens the door for innovative approaches to account for operational variables in power curves.M.S.Includes bibliographical reference