Is it returning too hot? Time series segmentation and feature clustering of end-user substation faults in district heating systems

This study explores the challenges and advancements in collecting ground-truth data to enhance fault diagnosis models for district heating systems. Initiated by the need to address limitations in previous data collections, this research leverages an enriched dataset from a Danish district heating utility to identify faults in household substations. Despite some inaccurate fault categorizations, complex fault patterns, and truncated measurements, the analysis of 50 detailed cases out of 127 fault reports reveals that, while return temperature reliably indicates faults, energy usage patterns do not. By employing self-organizing maps combined with k-means clustering, fault symptoms and patterns were categorized adequately, demonstrating the utility of high-dimensional data clustering in fault diagnosis. Additionally, an algorithm using time series decomposition is suggested to identify extreme and subtle anomalies, enhancing fault detection capabilities. The paper concludes that these methodologies significantly improve the accuracy and dependability of fault diagnostics in district heating systems, paving the way for more efficient operational management. • Self-organizing maps and k-means for fault clustering in heating systems. • 50 cases were analyzed, revealing distinct features related to a fault occurrence. • Demonstrating time series decomposition for automated anomaly identification. • Highlighting challenges in data truncation and fault diagnosis accuracy. • Proving the effectiveness of data segmentation in detecting various fault anomalies.