Anomaly detection for solar thermal parabolic trough power plants with artificial intelligence

Abstract

Defects and faults in parabolic trough power plants often lead to lower energy production. The automated detection of such anomalies could reduce downtimes and increase efficiency. The anomaly detection is based on sensor data which are measured anyway for the operation of the plant. The measured data are high-dimensional data in a spatio-temporal context. In the first step, the efficient preprocessing and visualization of the solar field and weather data is realized. From the existing sensor data useful features are extracted, which build the input for further approaches of anomaly detection models. Regarding the given problem, approaches that make use of methods of artificial intelligence are presented. One approach is pursued. The basic idea of this approach is the consideration of multivariate time series on loop level, which are segmented and clustered. For the segmentation of the time series three different segmentation methods based on the bottom-up or sliding-window principle are compared. For clustering, the density-based DBSCAN algorithm is used. The results of the anomaly detection show that about 93 % of the detected time series segments behave unusually from the data point of view due to manual interventions in the operation of the power plant. Suggestions for an improvement are given to reduce the number of false detections. Furthermore, weaknesses of the implemented approach are pointed out.