ARCHIVES

An adaptive Graph-Attention (GAT) spatial layer has previously been shown, in earlier unpublished work by the author, to outperform a fixed, distance-based spatial weighting scheme for hyperlocal PM2.5 forecasting on a 62-station monitoring network in Gurugram, India, under a controlled, capacity-matched comparison. Like nearly every graph-based air quality forecasting study in the reviewed literature, that result was established on a single city, leaving open whether the advantage is a property of the method or an artefact of that network's density or geography. This paper reports an independent cross-city validation: the identical fixed-weight-versus- adaptive-GAT comparison, matched in parameter count and trained under an unmodified protocol, is applied to a separately collected dataset from Mumbai, India. Of 29 candidate CPCB/MPCB/BMC stations (hourly data, 2025), 19 were retained after a data-quality screen; wind speed was unavailable, so the feature set was reduced from seven inputs to six, applied identically to both models. Across three seeds and four forecast horizons (1, 6, 12, 24 hours — 12 seed-horizon comparisons), the adaptive model achieved a lower RMSE than the fixed-weight model in 10 of 12 cases; a paired Wilcoxon signed-rank test rejects the null hypothesis of no systematic difference (W = 3, p =0.0024, two-sided). The two exceptions were both small and both at the 24-hour horizon. These results provide direct empirical evidence, to our knowledge not previously reported for this comparison, that the advantage of adaptive over fixed spatial weighting is not specific to the Gurugram network, while surfacing genuine dataset differences a future cross-city comparison should account for.