Application of Hybrid ML Models for Pipeline Failure Prediction

This paper presents a two-level hybrid machine learning pipeline for predicting annual failure risk in urban district heating and water supply pipeline sections. The dataset comprises 8,243 unique pipeline assets (Sys) aggregated into 1,920 sections, with failure statistics extracted for 2019–2025. At the first level, ensemble ML models (Gradient Boosting, Random Forests, ExtraTrees) assess failure risk of individual Sys assets using static engineering parameters and historical failure variables. At the second level, Sys-level scores are aggregated to sections using max pooling. On the 2023 test set (490 failure-prone sections out of 1,920), the best risk-detection mode achieved ROC-AUC of 0.8539, Recall of 0.9388, and Balanced Accuracy of 0.8222, correctly identifying 460 out of 490 pre-failure sections one year in advance. In the TOP-20 highest-risk sections, 17 were confirmed pre-failure (Precision@20 = 0.85). A selective high-confidence mode achieved 0.8412 accuracy at 69.22% automated coverage.