基于物理引导的集输管道内腐蚀速率预测及可解释性分析  

Physics-guided Prediction of Corrosion Rate Inside Gathering and Transportation Pipelines and Explanatory Analysis

作　　者：陈潜 黄伟[1] 张昌会 管奥成 张宸 叶晓芃 CHEN Qian;HUANG Wei;ZHANG Changhui;GUAN Aocheng;ZHANG Chen;YE Xiaopeng(School of Petroleum Engineering,Yangtze University,Wuhan 430100,China;Central Sichuan Oil and Gas Mine,PetroChina Southwest Oil and Gas Field Company,Suining 629000,China;Chongqing Division,PetroChina Southwest Oil&Gasfield Company,Chongqing 400707,China;Central and Northern Sichuan Gas Production Management Office of Southwest Oil and Gas,Field Company,Suining 629000,China)

机构地区：[1]长江大学石油工程学院,武汉430100 [2]中国石油西南油气田分公司川中油气矿,遂宁629000 [3]中国石油西南油气田分公司重庆气矿,重庆400707 [4]中国西南油气田分公司川中北部采气管理处,遂宁629000

出　　处：《中国腐蚀与防护学报》2025年第3期720-730,共11页Journal of Chinese Society For Corrosion and Protection

基　　金：中国石油西南油气田公司博士后基金(20220305-18)。

摘　　要：由于恶劣的服役环境,集输管道的腐蚀问题愈发严重,亟需提出一种准确的模型来预测集输管道的内腐蚀速率。本文首先提出了一种基于物理引导神经网络(PGNN)与改进粒子群算法(IPSO)相结合的管道腐蚀预测方法。通过分析电化学腐蚀机理,总结不同腐蚀因素变化对腐蚀速率的普适性影响规律。然后基于这些普适性规律提出损失函数表征方法,构建物理引导的内腐蚀速率预测模型。使用改进粒子群算法优化模型的超参数。最后结合电化学腐蚀机理和部分依赖图、SHAP算法对预测模型进行可解释性分析。结果表明,PGNN-IPSO方法可以避免模型学习到与物理相悖的错误规律,提升模型预测准确度。该研究对集输管道的腐蚀防护、可靠性评估和维修决策具有重要意义。Due to the harsh operating environment,the corrosion problem of gathering and transportation pipelines has become increasingly severe,necessitating the development of an accurate model to predict the internal corrosion rate of such pipelines.Herein,a pipeline corrosion prediction method that combines physics-guided neural networks(PGNN)with an improved particle swarm optimization(IPSO)algorithm is proposed.By analyzing the electrochemical corrosion mechanism,the universal impact of variations in different corrosion factors on the corrosion rate is summarized.Subsequently,based on these universal laws,a loss function characterization method is proposed to construct a physics-guided internal corrosion rate prediction model.The model's hyperparameters are optimized using a particle swarm algorithm improved through reverse learning,nonlinear weight adjustment mechanisms,and cosine algorithm enhancements.Finally,the predictive model is subjected to interpretability analysis using electrochemical corrosion mechanisms,partial dependence plots,and the SHAP algorithm.The accuracy of the model is evaluated using four metrics:mean absolute error(MAE),root mean square error(RMSE),mean absolute percentage error(MAPE),and R-squared(R2),while its reliability is assessed based on its consistency with corrosion mechanisms.Results demonstrate that the PGNN-IPSO method can avoid learning erroneous patterns contradicting physical principles,thereby enhancing prediction accuracy.This research holds significant implications for corrosion protection,reliability assessment,and maintenance decision-making regarding gathering and transportation pipelines.

关 键 词：集输管道 内腐蚀 物理引导神经网络 可解释性 改进粒子群算法 SHAP算法 部分依赖图 

分 类 号：X397[环境科学与工程—环境工程]