Improving wind power forecasts : integrating adaptive histogram of oriented gradients and multiple numerical weather predictions

Abstract

In this research, alternative representations of the grid based Numerical Weather Prediction (NWP) models are proposed for wind power forecasting purposes. Wind speed is the major indicator for the power generation. However, wind direction has nonlinear effects coming from the wind’s dynamic nature. With the traditional representations, the continuous and cyclic behavior of the wind direction is not perceived by the learners. To address these problems, Standard and Supervised transformation methods based on the Histogram of Oriented Gradients (HOG) of the NWP models are proposed. Our experiments on forty-seven wind farms show that the Standard HOG transformation on the naive features of multiple NWP models from distinct locations provides superior performance in linear learners. Additionally, Supervised HOG transformation is proposed with the utilization of tree based clustering methods. According to the results of the experiments with combined representations, linear learning methods outperform the tree based learners in wind power forecasting. This result indicates that the nonlinearity deriving from the wind direction’s circular behavior is represented in as suitable form for the linear learners.