A1288
Title: Short-term probabilistic energy load forecasting with GAMLSS framework
Authors: Katarzyna Chec - Wroclaw University of Science and Technology (Poland) [presenting]
Bartosz Uniejewski - Wroclaw university of Science and Technology (Poland)
Rafal Weron - Wroclaw University of Science and Technology (Poland)
Abstract: Electricity demand forecasts are among the key determinants of electricity prices and are therefore widely used in energy price forecasting models. In practice, decision makers often rely on forecasts provided by transmission system operators (TSOs), which tend to exhibit limited accuracy and systematic bias. Recent research has shown that these forecasts can be significantly improved using simple autoregressive models with day-ahead information, yielding notable reductions in forecast errors. However, the focus is solely on point forecasts, without addressing probabilistic approaches. The aim is to fill this gap by investigating the applications of probabilistic forecasting methods for improving TSO electricity demand predictions. Both post-processing of point forecasts (i.e., building predictive distributions based on point predictions) and direct distributional forecasting are considered, with a focus on generalized additive models for location, scale, and shape (GAMLSS). The analysis covers nine years of data (2016-2024) from three major European electricity markets. Forecast accuracy is evaluated using the pinball score and statistical significance is verified with the Diebold-Mariano test. It is found that GAMLSS outperforms post-processing approaches as well as quantile regression, and the use of corrected load forecasts yields superior performance over raw TSO predictions when employed as inputs to probabilistic models.
