Fault-Detection (FD) is essential to ensure the performance of solar thermal systems. However, manually analyzing the system can be time-consuming, error-prone, and requires extensive domain knowledge. On the other hand, existing FD algorithms are often too complicated to set up, limited to specific system layouts, or have only limited fault coverage. Hence, a new FD algorithm called Fault-Detective is presented in this paper, which is purely data-driven and can be applied to a wide range of system layouts with minimal configuration effort. It automatically identifies correlated sensors and models their behavior using Random-Forest-Regression. Faults are then detected by comparing predicted and measured values.

The algorithm is tested using data from three large-scale solar thermal systems to evaluate its applicability and performance. The results are compared to manual fault detection performed by a domain expert. The evaluation shows that Fault-Detective can successfully identify correlated sensors and model their behavior well, resulting in coefficient-of-determination scores between R²=0.91 and R²=1.00. In addition, all faults detected by the domain experts were correctly spotted by Fault-Detective. The algorithm even identified some faults that the experts missed. However, the use of Fault-Detective is limited by the low precision score of 30% when monitoring temperature sensors. The reason for this is a high number of false alarms raised due to anomalies (e.g., consecutive days with bad weather) instead of faults. Nevertheless, the algorithm shows promising results for monitoring the thermal power of the systems, with an average precision score of 91%.

Read the full article here.