Mean-deviation-driven synthetic data generation under data scarcity for energy demand time series

Deep learning in residential energy management is often limited by scarce, high-quality data. We propose a lightweight, interpretable method for synthetic electricity demand generation that combines statistical aggregation with deviation modeling, which requires no intensive training. We systematically compare it with Bayesian sampling, bootstrapping, conditional variational autoencoders, and conditional Wasserstein generative adversarial networks across five residential datasets at 1- and 15-minute resolutions, evaluating fidelity, diversity, utility, efficiency, and forecasting performance under varying data availability, synthetic-to-real ratios, and insertion strategies. Results show that one week of real data suffices to generate synthetic sequences that improve forecasting, with the largest improvements in data-scarce settings, particularly when augmentation roughly doubles the training set, and diminishing gains as more real data become available. Temporal coherence proved important, while temporal positioning had negligible impact. The proposed method provides an efficient, robust, and practical framework for improving short-term forecasting through synthetic data augmentation in real-world residential energy systems.