Comparative Analysis of Regression Models for Predicting a Synthetic Corrosion Defect Severity Indicator in Pipelines

This study presents a comparative analysis of regression models for predicting a synthetic corrosion defect severity (CR) indicator in pipeline systems. The investigated dataset contains 10,292 observations, eight input features, and a synthetic target variable. Fifteen models and configurations were compared: Dummy Regressor (Dummy), Linear Regression (LR), Ridge Regression (Ridge), Lasso Regression (Lasso), Elastic Net (EN), second- and third-degree polynomial regression (Poly2 and Poly3), k-Nearest Neighbors (KNN), Support Vector Regression (SVR), Extra Trees Regressor (ETR), Gradient Boosting Regressor (GBR), Histogram-based Gradient Boosting Regressor (HGBR), AdaBoost Regressor (ABR), Multilayer Perceptron (MLP), and CatBoost Regressor (CBR). The data were split into an 80% development set and a prespecified 20% held-out internal test set. Model selection and hyperparameter tuning were performed without accessing the test set using repeated nested cross-validation, with five outer folds repeated five times and five inner folds. The lowest mean root mean square error (RMSE) in the outer cross-validation was achieved by second-degree polynomial regression, with an RMSE of 0.019557 and a standard deviation of 0.001796; the mean absolute error (MAE) was 0.003762, and the coefficient of determination (R2) was 0.721530. However, its advantage over SVR was not statistically significant after Holm correction (adjusted p = 0.0662). On the prespecified held-out internal test set, Poly2 achieved an RMSE of 0.021454, an MAE of 0.003967, and an R2 of 0.679304. In the highest CR decile, the error increased to an RMSE of 0.050622, the bias was −0.017468, and R2 decreased to −1.372106, indicating systematic underestimation of high CR values. Permutation importance analysis identified carbon dioxide (CO2) content, gas production rate, pressure, water production rate, basic sediment and water (BSW), and oil production rate as the most influential features. These findings are limited to a computational experiment conducted on synthetic data and require external validation using actual corrosion measurements.