Urban air pollution forecasting is a critical component in the design of sustainable and smart cities, as it directly influences public health policy, transportation regulation, and environmental resilience. This paper presents a hybrid deep learning (DL) framework that integrates graph convolutional networks (GCNs) with long short-term memory (LSTM) units to perform multivariate spatiotemporal predictions of six major air pollutants: PM2.5, PM10, SO₂, NO₂, CO, and O₃. Unlike conventional univariate or grid-based time series models, our approach leverages the underlying topological structure of sensor networks and the temporal dynamics across multiple pollutant variables. A spatial graph is constructed based on the geographic locations of twelve monitoring stations in Beijing, enhanced with pollutant-specific correlations to encode both proximity and functional similarity among the sensors. The proposed GCN-LSTM architecture was trained on over 35,000 hourly observations per station. It demonstrates robust forecasting capabilities, achieving a root mean square error (RMSE) as low as 0.0316 and an R² value of up to 0.87 across various pollutants. Comparative experiments confirm the superiority of the hybrid model over baseline architectures, such as standalone LSTM and GRU models. This emphasizes the effectiveness of spatiotemporal graph representation in capturing urban air pollution dynamics. This framework provides a scalable and real-time solution for air quality management, offering a valuable tool for policymakers and urban planners engaged in smart environmental governance.