A Physics-Informed Variational Autoencoder for Modeling Power Plant Thermal Systems

Data-driven models for complex thermal systems face two main challenges: a heavy dependence on high-quality training datasets and a “black-box” nature that makes it difficult to align model predictions with fundamental physical laws. To address these issues, this study introduces a novel physics-informed variational autoencoder (PI-VAE) framework for modeling thermal systems. The framework formalizes the mechanistic relationships among state parameters and establishes mathematical formulations for multi-level physical constraints. These constraints are integrated into the training loss function of the VAE as physical inconsistency losses, steering the model to comply with the system’s underlying physical principles. Additionally, a synthetic sample-generation strategy using latent variable sampling is introduced to improve the representation of physical constraints. The effectiveness of the proposed framework is validated through numerical simulations and an engineering case study. Simulation results indicate that as the complexity of embedded physical constraints increases, the test accuracy of the PI-VAE progressively improves, with R² increasing from 0.902 (standard VAE) to 0.976. In modeling a high-pressure feedwater heater system in a thermal power plant, the PI-VAE model achieves high prediction accuracy while maintaining physical consistency under previously unseen operating conditions, thereby demonstrating superior generalization capability and interpretability.