Simple vs Complex: Evaluating Deep Learning Architectures for Decomposition-Based Electrical Load Time Series Forecasting

Decomposition methods, such as Variational Mode Decomposition (VMD), convert complex, nonstationary time series into band-limited Intrinsic Mode Functions (IMFs), enabling the isolation of interpretable temporal components and the analysis of trends, seasonality, and oscillatory behav- ior. These techniques have gained increasing attention in the scientific community and are commonly employed as prepro- cessing steps in hybrid forecasting frameworks. To evaluate the performance of different categories of forecasting algo- rithms on band-limited components, this study investigates nine neural architectures, including recurrent, feedforward, basis expansion, and transformer models, applying them to both synthetic IMFs and VMD-decomposed real-world elec- trical load data. Results show that simple or lightweight mod- els, such as Multi-Layer Perceptrons (MLPs) and Temporal Convolutional Networks (TCNs), often match or outperform more complex transformers, indicating that architectural in- novations designed for raw time series do not necessarily improve IMF forecasting.