Any-quantile probabilistic forecasting of short-term electricity demand: Fusing uncertainties from diverse sources

Abstract

Power systems operate under significant uncertainty arising from diverse and dynamic factors such as fluctuating renewable energy generation, evolving consumption patterns, and complex market dynamics. Accurately forecasting electricity demand necessitates advanced methodologies capable of capturing these multifaceted uncertainties. Our work develops any-quantile probabilistic forecasting framework, which enables the generation of forecasts for arbitrary quantile levels at inference time using a single trained model. This constitutes a substantial methodological advancement over traditional quantile regression techniques, which typically require training a separate model for each quantile or limiting predictions to a fixed set of predefined quantile levels. We show that integrating this framework into state-of-the-art neural architectures, specifically ESRNN and N-BEATS, yields superior distributional forecasting performance in the context of short-term electricity demand. Additionally, we develop the general Bayesian theory of cross-learning and link its latent objects with the elements of our architectures, providing a Fusion theory foundation for cross-learning from multiple power systems.

Empirical validation utilizing a comprehensive dataset of hourly electricity demand from 35 European countries showcases the efficacy of our approach, demonstrating superior predictive performance and enhanced quantile forecasting accuracy.