Abstract
Accurate regional ozone (O3) forecasting remains challenging due to complex multi-pollutant chemical interactions and the inherent sparsity of real-world monitoring data. Existing spatiotemporal models often fail to explicitly decouple heterogeneous variables and rely on error-prone imputation methods to handle missing inputs. To address these issues, we propose a Multi-pollutant Interaction Spatiotemporal Network (MIST-Net), which treats individual environmental factors as independent heterogeneous nodes and employs a dual-graph architecture: an Intra-Station Graph to capture cross-variable dependencies, and an Inter-Station Graph for spatial dependencies. Additionally, an integrated dual-channel missingness indicator mechanism enables the model to learn robustly and directly from incomplete data including missing records and unstructured data. Extensive evaluations on two diverse datasets (Los Angeles and Madrid) demonstrate that MIST-Net achieves highly competitive performance. For 1-hour forecasting, MIST-Net reduced Root Mean Square Error (RMSE) by 25.08% over SARIMA and 18.43% over DCRNN on the LA dataset. Under an extreme 40% random data missing rate, MIST-Net exhibited robustness, limiting the increase in RMSE to within 13.0% for 12-hour forecasting. Furthermore, attention-based analyses provide an exploratory perspective on empirical inter-variable lag correlations and learned spatial correlation topologies. Ultimately, MIST-Net offers a robust deep learning paradigm for resilient air quality monitoring under sensor-failure scenarios and at newly deployed, unmonitored sites.