AirGPT: Spatio-temporal large language model for air quality prediction

Abstract

Air pollution poses a critical threat to public health, ecosystems, and climate stability worldwide. Accurate air quality prediction is essential for informed policy-making, health risk mitigation, and environmental management, enabling proactive responses to pollution events and long-term planning for sustainable urban development. Despite advances, deep learning models for air quality prediction still face three critical challenges: heavy reliance on abundant historical data, difficulty in effectively fusing diverse information sources, and a lack of interpretability. To address these issues, we propose AirGPT, a large language model framework for air quality prediction. AirGPT integrates a specialized spatio-temporal encoder with a novel spatio-temporal instruction-tuning paradigm, enabling it to efficiently model complex spatio-temporal dependencies and perform sophisticated data fusion. Furthermore, our Chain-of-Thought distillation mechanism allows the model to externalize its predictive reasoning in a transparent, human-readable format, thereby enhancing interpretability. Experimental results demonstrate that AirGPT achieves state-of-the-art accuracy on air quality prediction tasks, particularly in data-scarce and zero-shot scenarios. By integrating interpretable reasoning with transparent predictive outputs, AirGPT provides a robust and reliable framework to support informed environmental decision-making. Our source code is available at: https://anonymous.4open.science/r/AirGPT-6ACC