Modeling building behavior is important to enhance energy management and occupant comfort. Traditional simulation methods for district heating networks (DHNs) often struggle with scalability and computational efficiency. This paper explores diffusion models, a state-of-the-art deep learning approach, to generate conditioned multivariate time series data for DHNs, utilizing public building data and telemetry. Preliminary results suggest that diffusion models have the potential to capture statistical properties of time series, especially for durations shorter than 24 hours. While conditional generation did follow the real data in terms of mean trend and variance, challenges remain in accurately reproducing peaks and extreme values. We discuss potential improvements in embedding methods, evaluation, and model architectures to enhance robustness. Our findings highlight the promising use of diffusion models for DHN simulations.
