Generation of synthetic load profiles for different typologies of residential users through metadata-driven generative AI models

Generating synthetic yet realistic electrical load profiles for residential users is essential for energy scenario analysis, planning, and optimal design. While generative deep learning models have shown promise in producing synthetic load profiles, they often overlook contextual information, limiting their ability to capture the diversity of user typologies and behaviours. To address this, a Conditional Variational Autoencoder (CVAE) is proposed, integrating user metadata, historical consumption patterns from 35 real residential units (e.g., apartments, single-family houses), and exogenous factors to generate realistic daily electricity profiles for both consumers and prosumers. By capturing the probabilistic variability of residential electrical loads, the model accurately represents energy usage and production patterns while preserving macroscopic and long-term relationships. Its effectiveness was further demonstrated in a data-scarce context by simulating the operation of a small Renewable Energy Community (REC) in Italy, where it successfully captured load variability driven by occupancy patterns, appliance presence, and seasonality. The results outperformed those obtained with a conventional REC estimation approach, highlighting the model potential for improving energy planning and management.