Propagating uncertainty in tree-based load forecasts

This paper discusses the use of ensembles of regression trees as a straightforward but versatile methodology to generate short term (day-ahead) load forecasts for real data from the Global Energy Forecasting Competition 2014. Since temperature is a strong predictor of load, we investigate how forecast uncertainty in temperature can affect the performance of the prediction model. To this end, a singular value decomposition (SVD) based approach is harnessed to simulate noisy but realistic temperature profiles. Our results show that as long as uncertainty is not exceedingly large, it is worthwhile to include temperature forecasts as predictors.