A hierarchical classifying and two-step training strategy for detection and diagnosis of anormal temperature in district heating system

Anormal temperature data caused by various reasons such as sensor faults and operation faults, which has a negative influence on heat metering and operation regulation in district heating system (DHS). However, it is difficult to detect and diagnose anormal temperature among massive unlabeled operation data. Therefore, this paper proposes a novel hierarchical classifying and two-step training strategy to facilitate the anormal temperature detection and diagnosis task. Firstly, self-defined feature change rate of operation data like water temperature, flow rate, and valve opening are constructed as additional training features to capture the characteristics of anormal temperature conditions. Then, a hierarchical classifying method is proposed to detect anormal temperature data. Finally, a two-step training strategy which combines expert knowledge with support vector machine (SVM) to fulfill anormal temperature type diagnosis. The proposed strategy is applied to a typical DHS in cold region of China. A total of 10,920 anormal data are detected. Four anormal temperature conditions are diagnosed including offline sensor, inversely connected sensor, anormal operation of heat source, and shutdown of heat station. The diagnosis accuracy for the 4 kinds of anormal temperature conditions all reached over 98%.