HF-STGN: A robust and unified residual-enhanced spatio-temporal graph network for urban meteorological forecasting

With the rapid advancement of smart city infrastructures and the Meteorological Internet of Things (IoT), high-precision, microscale weather forecasting has become a pivotal enabler for refined urban governance. However, existing models generally lack generalizability. Most existing methods target isolated meteorological variables and lack a universal framework for complex multivariate data. Furthermore, their reliance on stationary mean regression significantly degrades performance when confronting non-stationary dynamics, particularly during extreme climatic events. To address these challenges, we propose the Heterogeneous Feature Residual Spatio-Temporal Graph Neural Network (HF-STGN), an end-to-end framework for heterogeneous signal decoupling and spatiotemporal reconstruction. First, a latent projection layer extracts orthogonal representations to adaptively fuse heterogeneous observations. Second, Residual Graph Blocks model spatial dependencies, creating a direct pathway for high-frequency signals. This mitigates over-smoothing and effectively preserves abrupt extreme values. Finally, a Gated Recurrent Unit decodes the nonlinear temporal patterns of latent features. Using high-density meteorological data from Yan'an, China, we performed extensive comparative experiments. These tests focused on temperature and atmospheric pressure, characterized as stationary and non-stationary fields, respectively. The results show that HF-STGN achieves superior performance over advanced spatiotemporal and attention-based architectures, lowering the temperature MSE to 0.0469 in standard scenarios. Robustness verification under representative extreme events further demonstrated that HF-STGN maintained low prediction errors during both severe cold-air outbreaks and anomalous low-pressure processes. This robust performance confirms the framework's generalizability for multi-element forecasting and its exceptional reliability in complex, non-stationary environments.