Energy structure analysis of the process heating system in crude glycerin refinery using a physics-guided dual-region echo state network for multirate steam modeling

Document Type

Article

Publication Date

2-1-2026

Abstract

Industrial steam consumption plays a critical role in ensuring energy efficiency and operational reliability in chemical processes. However, statistical and deep learning methods often struggle with irregular and multirate data collection, where critical process variables are sampled at different time intervals. Current modeling procedures generally consider only the multirate problem of input data fusion, while the multirate between target variables remains a critical problem in energy applications. To address this limitation, this study proposes a physics-guided dual-region echo state network (DR-ESN) for steam consumption prediction in a crude glycerin purification process. The model introduces a dual-reservoir structure that separates fast-sampling latent variables governed by the physical relationship of system balances from slower, nonlinear dynamics captured in a bulk reservoir. A buffered physics-guided loss function enforces physical feasibility while tolerating minor deviations due to noise and process fluctuations using the loss of latent variables. By benchmarking against a conventional deep learning model, the results show that the DR-ESN substantially outperforms baseline models, achieving an R2 of 0.9608, RMSE of 45.7342 kg/h, and MAPE of 0.8862 % on testing data. Beyond predictive accuracy, the framework provides interpretable insights into steam allocation across key process units (dryer, reboiler, and vacuum ejectors), enabling energy structure for scenario-specific diagnosis. The findings demonstrate that integrating physics-based constraints with reservoir computing not only improves prediction under multirate conditions but also enhances interpretability, offering a robust pathway for sustainable industrial energy management.

Keywords

Artificial intelligence, Echo state neural network, Glycerin refinery, Physics-guided model, Steam consumption

Publication Title

Applied Thermal Engineering

ISSN

1359-4311

DOI

10.1016/j.applthermaleng.2025.129365

Funders

Kasetsart University, Faculty of Engineering, Kasetsart University, Center of Excellence on Petrochemical and Materials Technology

Volume

286

Publisher

Elsevier Ltd

Share

COinS