The end-to-end smart lifetime prediction method for flexible thermal power plants: A case study on main steam pipe

• Published in: Energy (Oxford) Vol. 329; p. 136728
• Main Authors: Zhang, Wencan, Pan, Lei, Zhang, Junzheng, Chen, Ming
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.08.2025
• Subjects: Artificial neural networks; Creep lifetime prediction; Deep learning; Finite element method; SHAP; Finite element method; Deep learning; SHAP; Creep lifetime prediction; Artificial neural networks
• ISSN: 0360-5442
• DOI: 10.1016/j.energy.2025.136728

As thermal power plants have transformed in power grids from being the primary power source to being regulatory power source, deep and rapid load changes have become the norm. This frequently subjects thick-walled components under high temperature, such as boiler main steam pipes, to additional creep lifetime damage, posing security issues that cannot be ignored. Therefore, real-time monitoring of creep lifetime during operating conditions has become crucial. Since creep prediction for main steam pipes involves microscopic-scale calculations and is challenging to implement online, this paper proposes an online lifetime prediction method for main steam pipes that combines mechanism-based and data-driven modeling. Firstly, a finite element method is used to establish a mechanism-based model for pipe creep life. Secondly, a rapid creep lifetime deep learning prediction model based on stress prediction is introduced, where training data is generated from the mechanism-based model. This achieves an end-to-end intelligent prediction from operational data to real-time creep lifetime variations. The proposed method yields a root mean square error of 2.5 × 10−7 and an R-squared score of 0.925 on the test set, demonstrating good prediction accuracy. •Finite element model of steam pipe under different flexible conditions is established.•A feature selection method based on Pearson and Shapley additive explanations is proposed.•A new structure: Residual-Long short term memory Network is constructed.•A ‘two-step’ prediction method based on nearest neighbor interpolation is proposed.