Detecting Faults in Customer Substations to Reduce the High Return Temperature in District Heating

Sammanfattning

This study aimed to develop different methods of artificial intelligence to detect faults in district heating substations to reduce high return temperatures, in collaboration with Halmstad Energi och Miljö (HEM). Three approaches were applied: the statistical model Autoregressive Integrated Moving Average (ARIMA), the autoencoder-based approach using UnSupervised Anomaly Detection model (USAD), and the machine learning k-Nearest Neighbors algorithm (k-NN). Using unlabeled data from more than 4,000 buildings (2022–2025), faults were injected to evaluate the performance of the models. ARIMA used time series forecasting, achieving 51.1 % recall and 16.9 % precision. USAD demonstrated balanced performance, with 36.9 % recall and 36.8 % precision, while k-NN showed high recall, 77.4 %, and precision, 35.3 %. The study shows the precision-recall tradeoff, with method selection depending on how powerful the model is at detecting faults. Although k-NN shows high performance in detecting injected faults, minimizing the number of false alarms is required for operational deployment.