A0364
Title: Probabilistic forecasting of energy time series with diffusion models
Authors: Nicole Ludwig - University of Tübingen (Germany) [presenting]
Abstract: Energy systems are complex systems with multiple time series, such as available solar and wind power, interacting with external factors such as the weather and human behaviour. Probabilistic forecasting of these systems is crucial for managing the electricity network, acting upon imbalances and planning future energy system expansion. Recent advances in machine learning, such as transformers and diffusion models, show excellent capabilities in forecasting time series. However, in energy-related forecasting tasks, where seasonality and (future) covariates play an essential role, they rarely provide an additional benefit. The aim is to investigate how diffusion models can be enhanced to outperform simpler methods, especially when including complex future weather covariates and simultaneously forecasting multiple correlated time series. A particular focus is on the uncertainty estimates constructed by the different models, especially their marginal and conditional calibration and their ability to properly propagate the uncertainty from the weather input to the power output in the energy system. The base probabilistic forecasting performance regarding the calibration is compared, and statistical post-processing is assessed to see potential performance increases in the models post hoc.
