A diffusion-based generative framework for multi-energy scenario modeling in low-carbon building clusters

High-quality multivariate time-series data from building clusters are essential for energy system optimization, energy storage scheduling, and intelligent decision-making. Yet, their availability remains limited due to challenges in data acquisition, privacy concerns, and the inherent complexity of modeling multi-source energy interactions. While existing studies predominantly employ generative adversarial networks (GANs) to generate such data, GANs often suffer from unstable convergence during adversarial training, which can negatively affect the quality of generated samples. To address these issues, this study developed Spatio-Temporal Dependency Representation diffusion transformer (STDR-DiT)—a diffusion-based generative framework designed for synthesizing multi-source time-series data in building clusters. This model employs a dual-channel Transformer architecture to independently model temporal patterns and cross-dimensional dependencies. A STar Aggregate Redistribute (STAR) module is embedded in the temporal channel to enhance the capture of asynchronous correlations across energy modalities. To enhance controllability, a classifier-free guidance strategy is employed, which integrates conditioning variables into the generation process without the need for an external classifier. Experimental results show that STDR-DiT generates realistic and structurally coherent synthetic data, outperforming existing methods across multiple evaluation metrics. Ablation studies confirm the contributions of key architectural components. Finally, STDR-DiT is applied to generate uncertainty scenarios for day-ahead energy system scheduling under a stochastic optimization framework. Compared with robust optimization, the diffusion-based approach achieves lower carbon emissions and operational costs, highlighting its effectiveness for uncertainty-aware energy planning.