Abstract
Accurate and adaptive energy demand modeling in industrial systems is essential for improving operational efficiency, reducing costs, and supporting grid stability. However, industrial energy consumption is driven by complex, time-varying dynamics and latent factors, such as equipment degradation and operational regimes, which are difficult to monitor directly. Moreover, many existing data-driven approaches rely heavily on large-scale historical energy demand data, which limits their deployment in data-scarce and rapidly evolving industrial environments. This paper develops a probabilistic framework that models energy demand exclusively from exogenous operational and environmental inputs, capturing dynamic latent system behavior without requiring past energy consumption at inference time. Unlike conventional latent variable or recurrent forecasting models that rely on static embeddings, the proposed method infers evolving latent state trajectories using an Expectation-Maximization (EM) inspired learning procedure with sampling-based trajectory inference. To address data scarcity across heterogeneous industrial systems, we introduce a latent space transfer learning strategy in which dynamic latent representations serve as the transferable knowledge between systems, allowing efficient adaptation to new facilities by fine-tuning only part of the model. Numerical experiments demonstrate the effectiveness of the proposed framework in developing load models from exogenous variables and in leveraging transfer learning to construct individualized load models under limited data conditions.
[Display omitted]
•An exogenous probabilistic latent-variable framework is developed for energy demand forecasting.•Time-evolving latent state trajectories are learned to enable structured probabilistic modeling and uncertainty quantification.•A latent space transfer learning strategy is developed for rapid adaptation to new systems with limited data.•The transfer learning approach demonstrates competitive performance compared to baseline models.•Similarity-based source selection can improve transfer learning performance across heterogeneous domains.