Maskininlärningsbaserad avvikelseidentifiering i fjärrvärmebyggnader med hjälp av understationsdata

This thesis investigates the Machine Learning (ML) methods for District Heating System (DHS) anomaly detection using data collected from building substations. Nowadays, DHSes are widely deployed in Northern European countries such as Sweden and play a critical role in providing space heating for buildings. However, operational anomalies in DHS can degrade heating performance and potentially lead to system faults. Therefore, developing a reliable Anomaly Detection System (ADS) is of great importance for maintaining DHS efficiency and reliability. Given the limited availability of labeled data in real-world applications, this study employs unsupervised ML techniques for anomaly detection, distinguishing it from the more commonly used supervised approaches. The proposed ADS is applied to substations from three distinct DHS-powered building types. It integrates an Auto-Encoder (AE)-based detection algorithm with clustering methods, including Self-Organized Map (SOM) and K-means. The AE models are used to identify deviations from normal operational behavior, while the clustering algorithms categorize the detected anomalies. Experimental results demonstrate that the proposed models achieve effective anomaly detection across the selected building types. Moreover, the system can also identify not only severe anomalies but also frequent low-level irregularities, thereby supporting both fault prevention and long-term performance optimization of DHS operations.