Predicting Pipe Failures: A Machine Learning Approach to Asset Management in District Heating

Abstract

As district heating networks (DHNs) across Europe near or surpass their intended service life, the need for proactive, data-driven maintenance and replacement planning is increasing. This paper presents a predictive asset management framework based on machine learning (ML) to assess the fault vulnerability of pipes. To train the ML algorithm, a combined dataset of both geospatial and operational data is implemented. While pressure is often estimated using complex hydraulic simulations, this study proposes and evaluates a lightweight linear pressure approximation method based on expert knowledge. The ML model is applied to a real DHN in Denmark through collaboration with the utility company HOFOR. Due to a significant class imbalance in the data (~1:215), multiple oversampling (OS) strategies are investigated. Model performance is evaluated using fault capture versus length capture (FCLC) curves and area under the curve (AUC) metrics. The inclusion of pressure data is found to improve the AUC by 4.5%, and the model captures over 40% of failures by inspecting only 10% of the network. These results demonstrate the feasibility of using ML-based fault vulnerability prediction as a tool for predictive maintenance and asset replacement planning in DHNs.