A generative ozone concentration imputation and forecasting method based on a novel conditional diffusion model.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Ground-level ozone is a significant air pollutant in China, impacting environmental governance.
• Missing data in ozone time series hinders accurate analysis and prediction.
• Maintaining data integrity is crucial for effective policymaking.

Purpose of the Study:

• To develop a generative imputation and forecasting method for ozone time series data.
• To address the challenge of missing data in ozone pollution datasets.
• To enhance the accuracy and reliability of ozone concentration predictions.

Main Methods:

• Proposed SSSD-Transformer, a generative method combining structured state space diffusion (SSSD) and Transformer.
• Injected empirical knowledge as conditional information.
• Utilized SSSD and Transformer to capture temporal trends and detailed patterns in ozone pollution.

Main Results:

• SSSD-Transformer demonstrated accurate and robust imputations across various missingness scenarios.
• The model showed strong forecasting capabilities with stable performance as prediction horizon increased.
• Achieved excellent performance in both ozone concentration imputation and forecasting.

Conclusions:

• The SSSD-Transformer method offers a new perspective for ensuring data integrity in air quality monitoring.
• The approach enhances prediction credibility for ozone pollution.
• Contributes significantly to effective air pollution control strategies.