Small-Scale Leakage Detection in Water Distribution Networks Based on Parallel CNN–LSTM and Cross-Modal Attention Mechanism

Abstract Leakage in water distribution networks (WDNs) leads to substantial water loss and economic cost, and small-scale leakage persists because traditional detection methods identify it poorly. To address this challenge, a parallel convolutional neural network–long short-term memory (CNN-LSTM) with cross-modal attention (PCLA) model was proposed. A parallel architecture was employed to process spatial features from pressure-response heat maps and temporal features from pressure-response time series simultaneously, while a cross-modal attention mechanism was used to enhance the capture of critical leakage characteristics. The model was evaluated on eight WDNs of varying scales and topologies, including six distribution and two transmission systems, under simulated small-scale leakage scenarios (0.5%–2% of daily average system demand). On the L-town network, comparative experiments with five baseline deep learning methods showed that the PCLA achieved superior performance, with an F1 score of 0.9204 and false-positive rate of 0.0700. Cross-network experiments demonstrated that the PCLA generalized across all eight networks with F1 scores above 0.92. The model was further validated on six real-time leak scenarios from the BattLeDIM benchmark using 33 pressure sensors, achieving a 100% detection rate with 72.7% temporal coverage. The results indicate that the PCLA model provides a practical solution for small-scale leakage detection across diverse WDN configurations.