Abstract
<title>Abstract</title> <p>Extreme coastal water levels develop across several gauges during the same event, but the predictive value of a network depends on whether stations carry a shared atmospheric response or time-ordered information. We introduce CoastRelay, a recurrent graph framework that centres station states and represents lag-aligned gauge information and pairwise atmospheric forcing as separate inputs to a shared gated recurrent unit (GRU) forecast. Across the Thames–Southern North Sea, Charleston, and Gulf/Galveston networks, these inputs are combined in four otherwise identical graph conffgurations and evaluated for total water level and 95th-percentile (q95) mean absolute error over three random initialisations. The analysis reveals a regional complementarity regime: in the Thames–Southern North Sea, the joint representation adds upper-tail information beyond either single component, with an event-weighted q95 interaction of −0.0036 m (95% conffdence interval, [−0.0069 m, −0.0003 m]). This ffnding shows that cross-gauge coupling is most useful when lag-aligned water-level states and spatial atmospheric contrasts describe distinct parts of the same high-water event. CoastRelay provides a basis for choosing network information pathways when extending coastal water-level forecasting to new gauge systems.</p>