Intelligent Leak-State Identification Based on Instantaneous Pressure Characteristic Analysis and Prediction of Water Supply Networks

Abstract

The expansion of urban water supply networks has exacerbated leakage issues, necessitating prompt assessment and optimization of network operations. To address the limitations of traditional steady-state hydraulic analysis, this study proposes an integrated leak-state identification framework combining transient pressure prediction and pattern recognition. A hybrid model was developed to predict instantaneous pressure variations by integrating convolutional neural networks (CNNs), long short-term memory (LSTM), and an attention mechanism. Given laboratory normal and leakage conditions, the CNN-LSTM-Attention model achieved mean absolute error (MAE) values of 0.007815 ± 0.000283    kPa and 0.005345 ± 0.000324    kPa , root mean square error (RMSE) values of 0.008450 ± 0.000281    kPa and 0.005457 ± 0.000321    kPa , and coefficient of determination ( R 2 ) values of 0.95961±0.00265 and 0.93274±0.00776, respectively. The proposed model outperformed CNN, LSTM, bidirectional long short-term memory (BiLSTM), and CNN-LSTM models. Subsequently, an optimized support vector machine (SVM) model was constructed for leak-state identification using transient monitored pressure and the CNN-LSTM-Attention predictions, achieving an accuracy of 97.50%, precision of 100.00%, recall of 94.74%, specificity of 100.00%, and F1-score of 97.30% in laboratory testing. Furthermore, the proposed framework was validated using data from a real-world water supply network. Given field conditions, the pressure prediction model maintained an R 2 of 0.91895, whereas the optimized SVM achieved accuracy of 92.50% and an F1-score of 91.89%. Although model performance decreased due to operational variability and more complex hydraulic conditions, the framework still demonstrated feasible pressure prediction and leak-state identification capability given monitored field conditions.