Abstract: Pipeline corrosion is a critical safety hazard in oil and gas transportation systems.This is especially the case when multiple corrosive factors such as CO₂,H₂S,and Cl^- coexist,under which corrosion behavior becomes highly nonlinear and difficult to predict,posing challenges to traditional prediction methods based on experimental data or single-factor modeling.To address the limitations of existing prediction methods in terms of generalization ability,interpretability of key factors and real-time responsiveness,120 groups of high-frequency corrosion-monitoring data acquired by a field multisource sensing system integrating electrochemical,environmental,and compositional measurements were used.Spearman-Kendall correlation analysis,partial dependence plot(PDP) visualization,and an L_GL_1/2-regularized extreme learning machine(LF-ELM) were employed to establish a method for corrosion prediction and key-factor identification.The results show that inlet temperature has the most significant effect on corrosion rate,with a Spearman correlation coefficient of ρ=0.79,whereas CO₂ and H₂S exhibit threshold effects at concentrations of 33 mg/L and 5 682 mg/L respectively.In full-sample testing,the proposed method reduces the average absolute error (E_MAE) to 0.001 15 mm/a with R² reaching 0.864 5 and the number of redundant nodes reduced by 63%.In future time-series testing,it still maintains high stability,with E_MAE=0.001 28 mm/a.The model not only achieves high-accuracy prediction,but also has good engineering practicability,thus providing a theoretical basis and an intelligent tool for corrosion-risk management of oil and gas pipelines.

Key words: pipeline corrosion, L_GL_1/2 regularization, extreme learning machine, multi-sensor data, partial dependence plot