Probabilistic gradient boosting machines for GEFCom2014 wind forecasting

Abstract

This paper describes the probabilistic wind power forecasting method that was used to win the wind track of the Global Energy Forecasting Competition 2014 (GEFCom2014). Executing a consistent machine learning framework for fitting independent models for each wind zone and quantile allowed us to automate our process for the duration of the competition. We used gradient boosted machines (GBM) for multiple quantile regression, fitting each quantile and zone independently. Standard smoothing techniques were applied to the dominant input signal in order to adapt to forecast inaccuracies, and a cross-sectional approach was applied. We provide a technique for utilizing information about correlated wind farms efficiently, using a two-layer modeling approach. Our accuracy was consistent throughout the competition, meaning that it can be utilized for similar day-ahead wind forecasting tasks with minimal modeling effort.