Title: VMDNet: A Temporal Leakage-Free Variational Mode Decomposition Approach for Electricity Demand Forecasting

Abstract: Forecasting electricity demand is a complex task, primarily due to the strong multi-periodic nature of real-world demand series, which necessitates the effective modeling of recurring temporal patterns. By making such structures explicit, decomposition techniques can significantly enhance predictive accuracy. Variational Mode Decomposition (VMD), a robust signal-processing method known for its periodicity-aware decomposition capabilities, has gained increasing traction in recent years. However, prior research has frequently been hindered by issues related to information leakage and suboptimal hyperparameter tuning. To mitigate these challenges, we introduce VMDNet, a framework designed to preserve causality. This approach employs sample-wise VMD to prevent temporal leakage, utilizes frequency-aware embeddings to represent each decomposed mode, and leverages parallel temporal convolutional networks (TCNs) for decoding, thereby ensuring mode independence and efficient learning. Furthermore, VMDNet incorporates a bilevel optimization scheme inspired by Stackelberg games to effectively guide the selection of VMD’s two critical hyperparameters. Empirical evaluations across three widely utilized electricity demand datasets demonstrate that VMDNet consistently surpasses state-of-the-art baseline models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC