Machine Learning-Enhanced Leakage Detection in District Heating Networks: Integrating Improved Hydraulic Modeling and Signal Denoising for High-Accuracy Localization

Leakage detection of district heating network (DHT) can prevent environmental and financial losses. This work presents a novel method that not only detects the location of leaks, but also takes engineering into account. Considering the effects of orifice shape, pressure, and cavitation, the traditional orifice outflow expression was improved based on the unsteady hydraulic model. For the pressure signals measured at the heat station, the signals are denoised by Shannon entropy-based wavelet transform, after which the PCA-Feature Importance feature extraction method is applied. The features are used to identify leaks using three neural network algorithms (LinearModel, CNN, MultLiM) and three decision tree algorithms (RandomForest, XGBoost, LightGBM). The result shows that decision tree algorithm achieves a leak pipeline localization accuracy of 99%, with a localization length of 10 m for individual pipelines. Even with only 1/6 of the measurement points or a low sampling frequency of 5 Hz, this method can still achieve a leak pipeline localization accuracy of 95%. It also maintains a low false alarm rate (FAR) even for abnormal jumps greater than the leak fluctuation magnitude.