Machine learning-based leakage fault detection for district heating networks

This study proposes a machine learning-based leakage fault detection (LFD) method that locates leakage faults using variation rates of observation data from installed flowmeters and pressure sensors. In the proposed LFD method, a delayed alert triggering algorithm periodically checks the time series of make-up flow rate to identify if the network leaks. When a district heating system is in normal conditions, a hydraulic simulation model built based on the results of impedance identification is used to simulate all possible leakage faults in the network, and the resulting change rates of observation data constitute a leakage data set. An LFD model is then trained using the leakage data set. When the delayed alert triggering algorithm sends a leakage signal, a variation rate vector of observations of this leakage condition is collected. According to the vector, the LFD model can output the name of the leakage pipe. The proposed LFD method is validated in a case study, where the mean values of accuracy and macro-F1 score of the LFD results are 85.85% and 0.99786, respectively. The results show that the proposed LFD method provides a feasible and convenient solution for heat providers to timely and accurately detect leakage faults in networks.