摘要：Utilizing existing natural gas pipelines is an effective way for hydrogen energy transportation. However, hydrogen embrittlement can aggravate the material degradation, which can reduce the residual strength and pose a significant challenge to pipeline reliability. Consequently, in this paper, a residual strength prediction method is proposed for hydrogen-blended natural gas pipelines. Firstly, finite element analysis is utilized to investigate the failure pressure under different defect shapes, varying defect parameters and hydrogen concentrations. Secondly, a new feature space is constructed by integrating physically significant features tied to failure mechanism, and feature selection is conducted based on Pearson correlation coefficient. Then, a knowledge distillation model optimized by multi-objective improved hiking optimization algorithm (MOIHOA) is developed for residual strength prediction, the teacher-student framework is used to improve the computational speed and prediction performance. To overcome the challenge of limited hydrogen pipeline data, incremental learning mechanism is introduced for generalization of hydrogen-blended conditions. Finally, Shapley additive explanations (SHAP) is utilized to enhance the model interpretability. The results indicate that the proposed model can effectively predict the residual strength under hydrogen-blended conditions. This study provides a theoretical basis for reliability evaluation of hydrogen-blended natural gas pipelines.

关键字：Hydrogen-blended natural gas pipeline; Residual strength prediction; Knowledge distillation; Incremental learning; Model interpretability

ISSN号：0360-5442

卷、期、页：卷341

发表日期：2025-12-30

期刊分区（SCI为中科院分区）：一区

收录情况：SCI（科学引文索引印刷版）,EI（工程索引）,SCIE（科学引文索引网络版）

发表期刊名称：ENERGY

参与作者：苗兴园

通讯作者：马英涵,孙祥龙

第一作者：赵弘

论文类型：期刊论文

论文概要：苗兴园,马英涵,孙祥龙,赵弘，Residual strength prediction of hydrogen-blended natural gas pipelines based on incremental knowledge distillation，ENERGY，2025，卷341

论文题目：Residual strength prediction of hydrogen-blended natural gas pipelines based on incremental knowledge distillation