A Cross-Level Consistency-Aware Diffusion Model of Hierarchical Imputation for Energy Prosumers

Motivated by the aggregation structure of prosumers from energy communities within a region, this paper proposes a cross-level consistency-aware diffusion model for solving the hierarchical imputation problem (CDHI), directly processing three-dimensional (3D) hierarchical time series. In CDHI, a 3D attention mechanism of hierarchical, temporal, and feature Transformer layers is developed to capture the aggregated dependency of hierarchical series. Based on mask modeling, we extend imputation target choice strategies of missing patterns, thus training CDHI in a self-supervised manner. For implementation, we formulate the aggregation structure of prosumers as the hierarchical consistency across aggregation levels in the denoising loss function of CDHI.