From data scarcity to scalability: Generating realistic building energy profiles with GANs

Access to high-resolution smart meter data is often restricted by privacy regulations, limiting the development of scalable energy system analytics. This study presents an open-source framework for generating year-long, hourly synthetic residential electricity profiles using a convolutional Wasserstein GAN combined with a similarity screening that reduces the risk of reproducing individual training households. Unlike most existing approaches that synthesize daily or weekly sequences, the method captures daily, weekly and seasonal dependencies jointly by reshaping annual load profiles into a 24 × 368 matrix. The synthetic data are validated with statistical metrics across four temporal resolutions, and the model is benchmarked against TimeGAN, a transformer-based GAN and a diffusion model under identical training conditions. The WGAN performs best on all fidelity, temporal, detectability, diversity and energy indicator metrics, and in the benchmark on heat pump households it reproduces the distributions of annual consumption and peak load as closely as two sets of real households match each other. To demonstrate practical utility, an XGBoost classifier is trained to identify heat pumps, electric vehicles and photovoltaic systems on synthetic datasets of increasing size and evaluated on real data from 1300 Belgian households. With sufficiently large synthetic training sets, the classifier outperforms a classifier trained on the real data for most classes, with the largest gain for the underrepresented EV class. The approach facilitates the sharing of lower-risk long-term consumption data and provides a reproducible, scalable pipeline to support analytics that need smart meter data like forecasting, flexibility analysis, or energy system planning.