While the air quality prediction has been extensively studied, existing methods often struggle with stability due to complex data and modeling challenges. Data heterogeneity from various city areas can lead to significant information loss during representation generation, which involves the data complexity. Additionally, capturing geospatial-temporal dependencies amidst multiple uncertain factors adds to the model complexity. To tackle these dual challenges, effectively disentangling mixed impact factors from heterogeneous geospatial-temporal data is crucial for generating robust representations. In this paper, we propose a novel geospatial-temporal disentangled representation learning (GT-DRL) method for accurate air quality prediction. Our deep predictive model not only disentangles mixed data dependencies into independent geospatial-temporal dimensions for uniform reconstruction, but also models complex coupling relationships from multi-dimensional uncertain factors. Evaluations on two real-world datasets demonstrate GT-DRL’s superiority over baseline methods, highlighting its robustness and efficiency in handling data and modeling complexities.