Dear Reader,

For the 58th issue of our heatbeat Research Newsletter, we take a look at the latest research in fault detection at substations in district heating networks.

The first paper Prioritisation of faults in district heating substations: Towards predictive maintenance and optimised operation from Guevara Bastidas et al. covers a ranking of faults and their severity in district heating substations. The second paper introduces a new methodology for time series clustering in the field of fault detection. The title is HEAT: Hierarchical-constrained Encoder-Assisted Time series clustering for fault detection in district heating substations from van Dreven et al.

The study titled "Prioritisation of faults in district heating substations: Towards predictive maintenance and optimised operation" introduces a novel methodology to enhance the reliability and efficiency of district heating systems through improved fault management. With the increasing complexity and number of substations, traditional maintenance approaches are insufficient. The authors adapt the established Failure Modes and Effects Analysis (FMEA) into a new framework called Operation and Maintenance- Failure Modes and Effects Analysis (O&M-FMEA), which incorporates not only fault occurrence and severity but also monitoring potential and maintenance capability. This results in a new metric—the Maintenance Priority Number (MPN)—which better reflects the relevance of faults for predictive maintenance and operational optimisation.

To apply this methodology, the researchers compiled a comprehensive list of 81 faults in district heating substations, drawing from literature and German industry expertise. These faults were evaluated by 13 German practitioners across four dimensions: occurrence, severity, monitoring potential and maintenance capability. The survey results were used to calculate MPNs, enabling a ranked prioritisation of faults. The study found that faults such as contamination of strainers, pump failures, and heat exchanger fouling scored highest in MPN, indicating their criticality for predictive maintenance. These faults are not only frequent and severe but also offer good potential for early detection and intervention.

The study identified five faults with high relevance for predictive maintenance in German district heating substations: contamination of strainers (both primary and secondary sides), failure of the heating circuit pump, failure of the domestic hot water storage charging pump and fouling of heat exchangers. These faults were found to be frequent, impactful, and detectable before failure, with good potential for preventive or deferred maintenance. Conversely, faults such as air in the piping system and defective components of the domestic hot water electric 3-way valve showed low monitoring and maintenance potential, requiring alternative organisational strategies. Additionally, faults unrelated to O&M—like incorrect control unit parameterisation and misplacement of outdoor temperature sensors—highlight the need for improved installation and commissioning practices. These insights provide actionable guidance for utilities and researchers aiming to enhance fault detection and maintenance strategies in district heating networks.

The second research paper introduces HEAT (Hierarchical-constrained Encoder-Assisted Time series clustering), a novel unsupervised method for fault detection in district heating (DH) substations. The method addresses the limitations of traditional global or threshold-based fault detection approaches, which often miss subtle anomalies and require labelled data. HEAT operates in two phases: first, it approximates the relative network topology using supply temperature profiles and constrained hierarchical clustering, the implementation of soft constraints enables direct influence on cluster formation to categorize behaviour in district heating transfer stations in a more targeted manner; second, it performs intra-cluster anomaly detection using modified z-scores based on the Median Absolute Deviation (MAD). The method integrates a convolutional autoencoder (CAE) for dimensionality reduction and applies soft constraints to enforce cluster size balance and incorporate domain knowledge.

The study validates HEAT using real-world data from 248 substations in China, demonstrating its ability to form meaningful and balanced clusters that reflect the operational structure of DH networks. Compared to standard clustering methods (e.g., k-means, spectral clustering, DBSCAN), HEAT consistently outperforms in terms of cluster cohesion and size uniformity. The use of soft constraints allows for flexible integration of expert knowledge without rigid enforcement, improving interpretability and adaptability. The method also proves computationally efficient for medium-sized networks and offers a visual, explainable clustering structure through dendrograms.

HEAT achieved a sensitivity of 74.1% and specificity of 95.5% in fault detection, outperforming global detection (14.8% sensitivity) and other baselines. It successfully identified both major faults (e.g., valve failures) and subtle anomalies (e.g., secondary leakages), while maintaining a low false positive rate. The method’s unsupervised nature and ability to operate without labelled data make it particularly suitable for real-world DH systems, where labelled datasets are scarce. The study concludes that HEAT enhances operational efficiency and energy sustainability and can be adapted to similar DH networks with manageable recalibration, though future work is needed to test its scalability and generalisability.

Further Information

As always, we recommend reading the article in full. In addition to this research newsletter and various blog posts, we have added a monthly feature update to our blog, summarizing important developments and new features in our heatbeat Digital Twin. You can find the latest entry at https://heatbeat.de/en/blog/76/ .

The next issue of our newsletter will be published on September 3rd, 2025.