A component-aware transfer learning framework for day-ahead probabilistic load forecasting in sustainable power systems

Abstract

The increasing penetration of renewable generation and the growing spatiotemporal heterogeneity of electricity demand make accurate day-ahead probabilistic load forecasting increasingly important for low-carbon dispatch, reliable grid operation, and sustainable energy management. In practice, however, forecasting models trained in one region or period often degrade when deployed in another, because load series exhibit component-level domain shifts, such as phase variations in seasonal patterns and amplitude changes in long-term trends. To address this challenge, this paper proposes a component-aware transfer learning framework for cross-domain probabilistic load forecasting. The proposed method first decomposes load sequences into seasonal and trend components to capture their distinct temporal characteristics. It then performs bi-domain feature interaction through cross-attention and applies differentiated distribution alignment to the decomposed components, using Maximum Mean Discrepancy for seasonal alignment and an attention-enhanced CORAL strategy for trend adaptation. A quantile regression head is further introduced to generate probabilistic forecasts for uncertainty-aware operation. Experiments on multi-region electricity demand data from the Australian National Electricity Market demonstrate that the proposed framework generally improves both point forecasting accuracy and probabilistic reliability under temporal and regional transfer settings. These results indicate its potential for data-efficient forecasting deployment in heterogeneous power systems and for supporting resilient and sustainable energy system operation.