Day-ahead power load curve forecasting based on multi-objective trade-off feature construction and selection
TL;DR
Experimental results show that the proposed deep learning forecasting method significantly improves forecasting performance and curve smoothness across multiple real-world load datasets.
Abstract
Against the backdrop of the rapid development of modern power systems, day-ahead load forecasting faces increasingly complex challenges. Current methods largely rely on a single evaluation criterion for feature selection, making it difficult to balance feature relevance and redundancy. Meanwhile, the introduction of attention mechanisms often lacks clear motivation and visual performance demonstrations, limiting their full potential and impeding improvements in prediction accuracy and curve smoothness. To address these issues, this paper proposes a deep learning forecasting method that integrates feature construction with multi-objective feature selection. First, an extended feature set combining time series structural features and external driving factors is constructed, and a fitness function based on the Non-dominated Sorting Genetic Algorithm II (NSGA-II) is designed to achieve a multi-objective optimization balancing predictive relevance and feature redundancy. Second, dynamic weight allocation is implemented through an attention mechanism to focus on critical time periods, with visualizations intuitively demonstrating its effectiveness in suppressing prediction “sharp oscillations” and enhancing trend perception. Experimental results show that the proposed method significantly improves forecasting performance and curve smoothness across multiple real-world load datasets.