Trend-Aware Multi-Task Learning for Short-Term Energy Forecasting

Researchers have proposed two new methods for improving short-term energy forecasting and multi-task model fusion, enhancing both numerical accuracy and trend prediction performance.

A trend-aware multi-task learning framework has been developed for short-term energy forecasting, achieving competitive numerical accuracy and improving trend prediction performance^[1]. This framework decomposes forecasting outputs into directional movements and deviation magnitudes, adopting a task-specific dual-stream architecture and incorporating an uncertainty-aware task weighting scheme to stabilize multi-task learning. Short-term energy forecasting is crucial for real-time operational decision-making, such as electricity market bidding and power system dispatch. Existing forecasting approaches typically formulate the problem purely as a regression task, limiting their ability to capture directional movements and trend consistency required for operational decisions. Meanwhile, a new method called Concrete Subspace Learning based Interference Elimination has been proposed to improve multi-task model fusion by identifying a common low-dimensional subspace and utilizing its shared information to track interference^[2]. This method models the problem as a bi-level optimization problem and introduces a meta-learning framework to find the Concrete subspace mask.