Quantifying uncertainty: Air quality forecasting based on dynamic spatial-temporal denoising diffusion probabilistic...

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Air pollution poses significant health risks, necessitating accurate prediction models.
• Existing models often neglect prediction uncertainties and dynamic spatial-temporal dependencies.
• Deep learning models show promise but require advancements in capturing complex relationships.

Purpose of the Study:

• To develop an advanced air quality prediction model that quantifies uncertainties.
• To address limitations in prior models regarding dynamic spatial interconnections.
• To improve both short-term and long-term air quality forecasting.

Main Methods:

• Proposed the Dynamic Spatial-Temporal Denoising Diffusion Probabilistic Model (DST-DDPM).
• Utilized a denoising diffusion model to discern prediction indeterminacy.
• Incorporated a dynamic context encoder for dynamic adjacency matrices and a spatial-temporal denoising model for dependencies.

Main Results:

• DST-DDPM demonstrated superior performance over baseline models on a Beijing dataset.
• Achieved 1.36% improvement in short-term and 11.62% in long-term predictions.
• Successfully quantified prediction uncertainties in a case study.

Conclusions:

• The DST-DDPM effectively captures dynamic spatial-temporal dependencies in air quality data.
• The model provides reliable air quality predictions with quantified uncertainties.
• This approach advances the field of environmental monitoring and public health protection.