Transfer learning with graph neural networks for pressure estimation in monitoring-limited water distribution...

Water distribution networks (WDNs) constitute essential urban infrastructure, yet their monitoring is hindered by limited monitoring conditions. Soft sensing methods have been applied to estimate pressure at unmonitored nodes using the latest deep learning models, however, they rely on large datasets from the same WDNs for training. There is a critical gap in pressure estimation of WDNs under realistic monitoring limitations. This study proposes a Graph Neural Network-based Semi-supervised Transfer Learning (GASTL) approach that estimates node pressures by transferring knowledge between source and target WDNs. GASTL integrates a Heterogeneous Graph Neural Network (HGNN) to extract informative node representations, employs learnable shift parameters for domain adaptation to align source and target distributions, and incorporates Graph Laplacian regularization to enhance spatial consistency and estimation accuracy. The approach is tested on multiple benchmark WDNs, including C-Town, L-Town, and Ky13, under varying sensor numbers and network topology scenarios, and compared against various baseline transfer learning methods. Experimental results demonstrate that GASTL achieves an R² of 0.911 and a Mean Absolute Percentage Error (MAPE) of 9.15 % in Same-topology (e.g., C-Town to C-Town) transfers. In Cross-topology (e.g., L-Town to C-Town) transfers, it attains the same R² of 0.911 and a MAPE of 9.43 %. Further, the study identifies sensor numbers and placement as key factors influencing transfer performance. Notably, the number and location of sensors in the target WDN significantly affect estimation accuracy, whereas topological variations have minimal impact, as they primarily result in shifts in data distribution rather than structural constraints. These findings highlight the potential of transfer learning to improve WDN pressure estimation, offering a scalable and efficient solution for real-world applications.