Challenges in detection of heating curve misconfiguration: An empirical study

District heating networks require well-adjusted heating curves to ensure efficient heating while maintaining indoor comfort. Misconfigured heating curves can lead to insufficient heating or energy waste, yet determining when an adjustment is needed remains challenging. This study evaluates two data-driven methods for identifying substations that require heating curve updates and demonstrates why, under current data and labeling conditions, they largely fail to provide reliable signals for heating curve misalignment. First, we use the Prophet model to analyze return temperature trends, assuming that persistent deviations indicate a mismatch between supply temperature and building demand. Second, we train a conditional autoencoder (CAE) as a normal behavior model to detect anomalies before customers report insufficient heat.
The results show that the Prophet model is only reliable for certain substations. Its performance varies widely and often fails to capture complex patterns. The CAE successfully detects technical hardware faults, such as safety valve malfunctions that cause abrupt physical inconsistencies. However, incorrect heating curve settings usually result in gradual drifts within normal system temperature ranges rather than sudden anomalies. Consequently, the CAE interprets the suboptimal settings as normal behavior.
Key challenges include a lack of preventive maintenance data, since models are trained primarily on customer reports rather than operator interventions. Furthermore, data gaps and substation-specific models that overfit reduce the robustness of network-wide detection. These findings suggest that identifying misaligned heating curves requires more than just detecting abrupt sensor changes and highlighting the need for datasets that incorporate operator-led interventions and richer labels.