Load forecasting under distribution shift: An online quantile ensembling approach

Reliable load forecasting is crucial for power system operations but remains challenging under frequent distribution shifts caused by evolving consumption patterns and external disruptions. While deterministic methods (DLF) generate point predictions and probabilistic methods (PLF) capture uncertainty, existing approaches fail to bridge these paradigms to utilize PLF’s distribution insights for improving DLF accuracy under shifting conditions. To address this gap, we propose Adaptive Online Quantile Ensembling, a novel framework that integrates probabilistic insights into deterministic forecasting for robust online adaptation. Our method features dynamic quantile ensembling with long-term and short-term weight decomposition for balancing stability and responsiveness, as well as a detect-then-adapt strategy for adaptive fast-and-slow learning based on real-time error monitoring. Extensive experiments on post-COVID load datasets demonstrate significant improvements in accuracy and responsiveness over baselines, particularly during abrupt and gradual distribution shifts. This work establishes an effective approach to leverage probabilistic information for accurate load forecasting in dynamic, non-stationary environments.