Impact of normalization techniques on synthetic load profile generation using Deep Generative Models

Synthetic load profiles are increasingly employed in power system studies as a cost-effective and privacy-preserving alternative to extensive smart meter deployments, with deep generative models (DGMs) showing promising results in capturing complex demand patterns. However, the impact of data normalization on their performance remains insufficiently explored. Using datasets from a university smart grid and industrial consumers in Germany, this work systematically evaluates five normalization techniques—Min-Max, Standard, Robust, Max-Abs, and Quantile—on four representative DGMs: Wasserstein GAN with gradient penalty (WGAN-GP), variational autoencoder (VAE), nonlinear independent component estimation (NICE), and denoising diffusion implicit models (DDIM). Additionally, this study introduces DDIM for synthetic load profile generation, providing a deterministic and faster sampling approach compared to traditional probabilistic denoising models. Results based on statistical and temporal metrics indicate that Max-Abs normalization consistently yields more accurate and stable synthetic profiles across all models and datasets, while Robust and Quantile methods often degrade essential distributional features. These findings highlight the critical role of normalization in developing effective synthetic data generation pipelines for power system applications.