Objective  China's oil and gas pipeline networks are expected to reach 24×10⁴ km by 2025. Pipeline transportation has become one of the key means of transportation in the country. However, these pipelines are vulnerable to corrosion caused by the surrounding soil environment and other factors, which shortens their life in service. To ensure the safe operation of buried pipelines, accurately predicting the degree of corrosion is crucial.

Methods  This paper presents a prediction model for the pitting depth of buried pipelines, guided by the corrosion mechanism and combining a Random Forest (RF) algorithm with a Multi-Objective Optimization process. The incorporation of knowledge about the pipeline corrosion mechanism enhances the interpretability of the machine-learning (ML) model. By building on the interaction mechanisms among characteristic variables, new variables were created to better reflect the influencing factors of the surrounding soil environment. The Gini coefficients in the Random Forest algorithm were used to evaluate the importance of all features in the new characteristic space through calculations. Additionally, a Hybrid Multi-Objective Grey Wolf Optimization (HMOGWO) algorithm was adopted to determine the optimal hyperparameters of the RF algorithm. This feature selection approach was integrated with the multiobjective optimization process, considering three comprehensive optimization objectives: the number of features, prediction accuracy, and model stability. Using a defined comprehensive evaluation index, a comparative analysis of the Pareto solution set was conducted to obtain the optimal combination of feature subsets and hyperparameters. The resulting feature subsets, which are both representative and optimized for performance, contribute to improvements in model stability and prediction accuracy.

Results  The designed model was validated using a pitting dataset of real-world buried pipelines. By leveraging the combination of the three-objective HMOGWO algorithm and the RF model, it significantly outperformed the three-objective MOGWO algorithm, the two-objective HMOGWO algorithm, the two-objective MOGWO algorithm, as well as both the single-objective GWO algorithm and the single-objective PSO algorithm in terms of prediction performance and stability.

Conclusion  The proposed model has proven effective in accurately predicting the maximum pitting depth of buried pipelines. It is more interpretable and accurate in pipeline corrosion prediction, guided by the corrosion mechanism. This model is shown to be valuable in prolonging the service life of pipelines, highlighting its significance for practical applications in the oil and gas transportation sector.