AirFlow: Context Preserving and Multi-Rate State Mo...

AirFlow: Context Preserving and Multi-Rate State Mo... | AI Research

Key Takeaways

AirFlow is a framework designed to improve air quality forecasting by addressing the distinct statistical behaviors of different pollutants.
Recent methods have improved spatial dependency learning and meteorological covariate modeling.
However, pollutant channels are still passed through the same normalization rule and temporal backbone, using a shared latent representation for channel-specific distributions and changes at different rates.
To address this limitation, we propose AirFlow, a pollutant-aware dual-stream framework that operates on station multivariate observations without additional graph propagation or predefined signal decomposition.
AirFlow also requires only 0.0483M parameters and 0.0215G FLOPs, achieving high forecasting accuracy with low computational overhead.

Paper AbstractExpand

Accurate air quality forecasting is essential for public health and urban environmental management, but remains challenging because pollutant channels differ in periodicity and distribution drift, while their concentration trajectories contain both multi-scale dependencies and rapid changes. Recent methods have improved spatial dependency learning and meteorological covariate modeling. However, pollutant channels are still passed through the same normalization rule and temporal backbone, using a shared latent representation for channel-specific distributions and changes at different rates. To address this limitation, we propose AirFlow, a pollutant-aware dual-stream framework that operates on station multivariate observations without additional graph propagation or predefined signal decomposition. Specifically, AirFlow designs two novel blocks: (1) a statistic-guided normalization routing mechanism that selects a normalization path for each pollutant according to its 24-hour autocorrelation and distribution drift; and (2) a hierarchical dual-stream state model that combines multi-scale state space propagation with learnable response coefficients, where gated bidirectional cross-attention exchanges information and adaptively fuses the resulting representations. Experiments on real-world data from multiple cities show that AirFlow achieves the best performance in 34 of 36 metrics comparisons, with reductions of up to 11.11% root mean square error over the state-of-the-art baseline. AirFlow also requires only 0.0483M parameters and 0.0215G FLOPs, achieving high forecasting accuracy with low computational overhead.

AirFlow is a framework designed to improve air quality forecasting by addressing the distinct statistical behaviors of different pollutants. Instead of applying a uniform processing method to all data, the researchers—Fan Yang, Nan Chen, Yijie Dong, Yuchen Zhang, and Wei Zhang—developed a system that adapts to the specific periodicity and distribution drift of individual pollutants, allowing for more accurate predictions of concentration trajectories.

Addressing Pollutant Heterogeneity

Standard forecasting models often treat all pollutants the same, using shared normalization and temporal processing. The authors note that this is problematic because pollutants like PM2.5 and PM10 exhibit abrupt, non-stationary fluctuations, while others like NO2 and O3 follow more predictable, periodic patterns. AirFlow solves this by using a "statistic-guided normalization routing mechanism." This system analyzes each pollutant's 24-hour autocorrelation and distribution drift to choose between two normalization paths: MinMax, which preserves absolute concentration levels for stable patterns, or RevIN, which reduces local shifts for more volatile data.

Dual-Stream State Modeling

To capture both long-term historical context and rapid changes in air quality, AirFlow uses a hierarchical dual-stream state model. This architecture consists of two main components:

Structural Memory Stream: Uses selective state-space propagation to maintain persistent historical context.
Transient Dynamics Stream: Uses learnable response coefficients to adjust how quickly the model reacts to new, short-term observations.
The framework uses "gated bidirectional cross-attention" to allow these two streams to exchange information. By conditioning the streams on one another, the model can adaptively fuse the representations to balance stable historical trends with immediate, sharp changes in pollution levels.

Performance and Efficiency

In experiments using real-world data from Beijing, Tianjin, and Hangzhou, AirFlow outperformed existing state-of-the-art baselines in 34 out of 36 metrics. The researchers reported that the model achieved up to an 11.11% reduction in root mean square error. Notably, the framework is designed to be computationally efficient, requiring only 0.0483M parameters and 0.0215G FLOPs. This allows for high forecasting accuracy without the need for complex, resource-heavy graph propagation techniques often found in other air quality models.

Considerations for Implementation

The authors emphasize that AirFlow is specifically designed for scenarios where detailed spatial data—such as synchronized wind fields or emission inventories—may be unavailable or coarsely sampled. By focusing on pollutant-aware temporal modeling rather than complex spatial graph construction, the model provides a way to maintain accuracy even when the environmental context is incomplete. The model's training process includes a trajectory-aware objective that penalizes errors in both point-wise concentration and the rate of change between forecasting horizons, ensuring the predicted pollution trends remain consistent.