VMDNet: Temporal Leakage-Free Variational Mode Decomposition for Electricity Demand Forecasting
Researchers have developed VMDNet, a new framework for electricity demand forecasting that addresses issues of temporal leakage and hyperparameter tuning in existing Variational Mode Decomposition (VMD) methods. The VMDNet framework uses sample-wise VMD to prevent temporal leakage, employs frequency-aware embeddings and temporal convolutional networks for efficient learning of decomposed modes, and incorporates a bilevel optimization scheme inspired by Stackelberg games to select VMD hyperparameters. Experimental results on three datasets demonstrate that VMDNet surpasses current state-of-the-art forecasting baselines. AI
IMPACT Introduces a novel method for time-series forecasting that could improve accuracy in energy demand predictions.