Comparison of different machine learning algorithms for prediction of wind evolution

Abstract

Wind evolution, i.e. the evolution of turbulence structures over time, has become an increasingly interesting topic in recent years. Part of this increasing interest is resulted from the development of lidar-assisted wind turbine control, which requires accurate prediction of wind evolution to avoid unnecessary or even harmful control actions. Moreover, study of wind evolution can deepen our understanding of turbulence characteristics and, for example, could make 4D stochastic wind field simulations possible by integrating wind evolution into 3D simulations. This study compares different machine learning algorithms for the prediction of wind evolution, including three variants of Gaussian process regression, regression tree, support vector machine regression, and shallow neural network. The comparison is done using lidar data measured by a nacelle-mounted lidar, and it focuses on the prediction accuracy and the computational time of models, to provide some insights of the trade-off between these factors. In terms of computational time, both the training time of the final model and the time for hyperparameter optimization are taken into account. Moreover, up-scaling of the computational time with data size is also investigated. It is found that Gaussian process regression provides high prediction accuracy but needs long computational time, increasing significantly with data size. For cases where long computational time might be a critical problem, shallow neural network could be a good alternative thanks to its high training efficiency, although its prediction accuracy is lower than Gaussian process regression. The other methods have no advantages in wind evolution prediction compared to these two methods.