Predicting node pressure accurately is of paramount importance for the management of water distribution networks (WDNs). Recent advances have highlighted the efficacy of graph neural networks (GNNs), tailored for data with inherent graph structures, in addressing this challenge. However, the performance of extant GNN-based approaches is constrained by their limited capacity to harness long-range dependencies within the network. To address this limitation, we introduce a novel long-range adaptive convolution network. Inspired by the graph kernel, our method possesses a broad receptive field, while the flexibility of information aggregation is enhanced through the attention mechanism. Additionally, we incorporate residuals specifically engineered for WDN applications to further refine our prediction accuracy. Our comprehensive evaluations on three real-world WDN datasets reveal that our method consistently surpasses existing benchmarks. We have made the code and experimental datasets publicly accessible via a GitHub repository ( https://github.com/Haldate-Yu/GAL-WDN ).