Autoencoder-based detection of incorrect parametrization in district heating substations

District heating networks rely on well-adjusted control parameters to deliver heat efficiently and maintain indoor comfort. One of the key parameters in this process is the heating curve, which determines the supply temperature as a function of outdoor conditions. If the heating curve is misaligned, it can lead to insufficient heating, customer complaints, and inefficient system operation. The aim of this thesis is to investigate whether data-driven methods can identify when the heating curve of a substation needs to be updated.
The study uses two approaches. First, the Prophet model is applied to normal operating periods to predict and decompose the return temperature into its trend, seasonal variation, regressor effects, and residuals. The idea behind this approach is that the return temperature, given the effect of outdoor temperature, should stabilize around a typical value of about 40°C when the heating curve is correctly adjusted. A persistent deviation in the trend may indicate that the supply temperature no longer matches the heating demand and that the heating curve may require updating. Second, a Conditional Autoencoder (CAE) is used to learn the normal behavior of substations and detect deviations that appear in the days leading up to customer-reported faults.
The results show that the Prophet model can decompose the return temperature reliably for only some substations. For others, the model fails to capture the underlying patterns, which limits its usefulness as a diagnostic tool. The CAE performs well in detecting many types of faults where the system behavior changes clearly, such as safety-valve malfunction or defective temperature control. However, faults related to incorrect heating curve parametrization do not produce strong or abrupt deviations in the supply or return temperatures. As a result, the CAE reconstructs these signals accurately and no deviations from normal behavior are detected. Consequently, the model does not detect when the heating curve must be updated. An analysis of the two heating curve update cases confirms that the effect of a misconfigured heating curve appears as a gradual drift rather than as a sudden abnormality in the monitored features.