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

In recent years, the generation of synthetic building load profiles has gained significant importance across multiple domains. These profiles are critical for applications such as energy simulation and modeling, where the use of realistic consumption data enhances accuracy, facilitates the identification of reference typological patterns, and supports the planning, design, and management of building clusters.

Deep learning techniques, including Generative Adversarial Networks (GANs) and Variational AutoEncoders (VAEs), have emerged as promising tools for the realistic generation of synthetic data. However, the load profiles generated by these models often lack the integration of crucial building metadata—such as user type, installed energy systems, and occupant numbers—as well as climatic conditions like temperature and irradiance. To overcome this limitation, advanced approaches such as Conditional GANs (CGANs) and Conditional VAEs (CVAEs) have been developed, enabling the integration of building metadata and exogenous conditions to generate more realistic synthetic load profiles.

This study addresses this challenge by proposing a deep learning-based methodology for generating synthetic load profiles for various types of residential users, including consumers, prosumers, and prosumers with energy storage systems. The approach leverages building and user metadata, features extracted from historical load time series, and exogenous conditions to develop a generative model capable of producing accurate daily load profiles for buildings. The model employed is a Conditional Variational AutoEncoder (CVAE), which incorporates covariate information to enhance the generation of synthetic data.

The proposed methodology is validated on a number of real buildings located in Italy, which are part of a Renewable Energy Community (REC). These buildings represent a diverse range of users, with available historical time series data on grid exchanges and detailed metadata regarding user type, number of occupants, rated power, and energy generation systems. By predicting energy consumption patterns based on user metadata and external conditions, the proposed tool generates load profiles that are dynamic and adaptable to varying climatic conditions and user types.

This capability not only provides valuable insights into the dynamics of the energy community when different user types could be added or removed but also supports its planning and design. Additionally, it facilitates active management through targeted user feedback and demand response (DR) programs, thereby enhancing the overall efficiency and resilience of the community.