GEFCom2014: Probabilistic solar and wind power forecasting using a generalized additive tree ensemble approach

We investigate the probabilistic forecasting of solar and wind power generation in connection with the Global Energy Forecasting Competition 2014. We use a voted ensemble of a quantile regression forest model and a stacked random forest-gradient boosting decision tree model to predict the probability distribution. The raw probabilities thus obtained need to be post-processed using isotonic regression in order to conform to the monotonic-increase attribute of probability distributions. The results show a great performance in terms of the weighted pinball loss, with the model achieving second place on the final competition leaderboard.