Oblique random forest algorithm using lasso regression for wind power forecasting

Abstract

With the increasing trend towards the use of renewable energy sources, wind power has been the subject of many researches. Wind power has stochastic nature due to uncertainties in atmospheric conditions, especially in wind speed, which makes it hard to forecast accurately. To solve the problem, statistical methods using Numerical Weather Prediction (NWP) models as inputs are proposed in the literature. Random Forest is a statistical model frequently used in wind power forecasting with proven success. Random Forest ensembles decision trees that partition the feature space over a single variable at each node. However, partitions based on a single vari able may fail to provide a proper distinction. Thus, oblique decision tree algorithms evaluating the partitions over linear combinations of variables are proposed in the lit erature, especially on classification problems. There are a limited number of studies in the literature on oblique decision tree-based methods applied in time series regression problems. This thesis proposes a novel strategy to be applied in regional wind power fore casting tasks that ensembles oblique decision trees. The proposed method is compared with its univariate counterparts in three wind power forecasting tasks. Computational results show that the proposed method performs better on all tasks.