<p>Air pollution remains one of the major environmental challenges in urban regions, with significant impacts on public health. Although numerous studies have addressed its prediction using time series analysis, many have overlooked spatial correlations between monitoring stations, which limits prediction accuracy. To address this limitation, we propose a general and novel approach to forecast a wide range of atmospheric pollutants using historical data from a network of air quality monitoring stations distributed across Spain, covering the period 2010–2020. Our method combines pollutant concentrations with meteorological variables and is based on a spatio-temporal graph architecture that integrates message-passing and spectral analysis techniques. One of the key innovations is the use of Singular Value Decomposition (SVD) applied to the adjacency matrix, enabling the capture of global structural properties of the network and improving the ability to learn long-range dependencies. The proposed model includes a Multi-Layer Perceptron (MLP) for adaptive feature fusion, as well as hierarchical message passing with a trainable aggregation mechanism. The effectiveness of the approach has been validated by comparison with several current deep learning models, yielding significantly superior results for all contaminants. This research provides a methodological contribution by combining graph spectral theory with spatio-temporal learning, offering a robust and generalizable framework for air quality prediction. Additionally, the model has potential applications in real-time environmental alert systems and public health surveillance platforms.</p>