An Inductive Spatio-Temporal Graph Network for Fine-Resolution Urban Weather Nowcasting Integrating Multi-Source Data and Physical Constraints

Fine-resolution urban weather nowcasting is crucial for urban resilience, yet it is fundamentally limited by the sparse and irregular distribution of monitoring stations. To overcome this, we introduce an inductive, physics-informed spatio-temporal graph network that transforms discrete sensor data into a continuous, on-demand forecast field. Our framework uniquely synergizes multi-source data: point-scale station observations, grid-scale numerical weather predictions, and high-resolution urban morphological features. The model core is a novel encoder-decoder architecture designed for deep feature extraction. A hybrid temporal encoder captures complex weather dynamics, while a multi-graph attention mechanism learns heterogeneous spatial interactions based on physical similarity (e.g., thermal or wind-driven connections), moving beyond simple geographic proximity. These multi-faceted features are then fused via a subsequent attention layer. Critically, we enforce physical consistency by integrating a thermodynamics-aware loss function, which ensures physics-informed predictions of key variables like temperature and humidity. Evaluated on a comprehensive dataset from Wuhan, China, our model demonstrated high accuracy and strong correlation with observational data for 6-hour ahead nowcasting. Its inductive design is a key advantage, enabling reliable predictions for arbitrary, unmonitored locations by leveraging their local morphological context. This work presents a scalable and robust framework for generating physically plausible, high-resolution urban weather intelligence, essential for proactive applications in energy systems, public safety, and climate-adaptive urban planning.