Due to its important role in public health and environmental protection, Air Quality (AQ) prediction has drawn increasing attention from data mining and machine learning communities. Despite the great progress that has been made, AQ prediction still faces several challenges: (1) Implicitly modeling spatial dependencies disregards location relationships between sparse monitoring stations. (2) The misalignment of the stations of the different types hinders the fusion of air quality and meteorological features. (3) The intricate multi-periodicity temporal patterns in AQ data are ignored. To tackle these challenges, in this paper, we propose Spatiotemporal-Aware Fusion Networks (STAFNet) for accurate AQ prediction. Specifically, we first design the Spatial-Aware Graph Neural Network to explicitly incorporate spatial information into representation learning. Then, we develop a Cross-Graph Fusion Attention to fuse multimodal features distributed on unaligned graphs. Moreover, we capture the multi-periodicity features and model temporal dependencies with TimesNet. Extensive experiments on three real-world datasets show that STAFNet achieves the best performance compared with ten baselines (3.3% \( \sim \) 5.0% on 25–48 hours future predictions). Our source code is available at https://github.com/dylan-yin/STAFNet_Paddle .