Probabilistic online learning framework for short-term wind power forecasting using ensemble bagging regression model

Abstract

The increasing penetration of renewable energy sources, with a notable focus on wind power, within modern electricity grids requires computationally efficient and burden-free short-term wind power forecasting models. Traditional models generating prediction intervals are trained offline and thus deployed for prediction purposes. However, this approach cannot obtain interval forecasts from the most recent wind power observations. In contrast, combining multiple regression models through ensemble learning is recognised as a successful method for improving forecasting performance. By utilising the most recent observations and exploiting the strengths of multiple regression models, this article investigates an Online Ensemble Bagging Regression (OEBR) model for generating prediction intervals. Online gradient descent based optimisation algorithms capable of adaptive-depth calculation from a stream of observations are used here to address the problems with traditional batch learning frameworks. The proposed online learning framework is evaluated against other online learning frameworks using publicly accessible datasets. The results show the proposed model competes with the compared models regarding probabilistic metrics and energy estimations and outperforms computational time requirements for the same number of observations.