Urban air quality modelling aims at inferring unknown pollution concentrations at specific urban locations. Physical methods derive the partial differential equations (PDEs) that mathematically define the laws of motion, albeit using computationally intense algorithms. In contrast, deep (generative) models, such as variational autoencoders, provide high performance by addressing the task as a data generation problem. Yet, physics knowledge and the spatio-temporal data correlations are not exploited by these deep learning models. In this work, we propose a physics-guided variational graph autoencoder whose graph convolutional operator is derived from the PDE defining the convection-diffusion physical process. We compare against statistical and deep learning approaches on two air quality datasets and report superior performance.