Probabilistic Deep Learning to Quantify Uncertainty in Air Quality Forecasting.

Area of Science:

• Environmental Science
• Computer Science
• Data Science

Background:

• Data-driven air quality forecasts have improved short-term predictions.
• Current models often lack reliable quantification of forecast uncertainty.
• Probabilistic deep learning offers tools for uncertainty estimation, but lacks empirical application in air quality forecasting.

Purpose of the Study:

• To apply and compare state-of-the-art uncertainty quantification techniques in air quality forecasting.
• To evaluate the performance, reliability, and applicability of probabilistic models for air quality predictions.
• To investigate methods for improving uncertainty quantification, including adversarial training and exploiting data correlations.

Main Methods:

• Trained probabilistic deep learning models, including Bayesian neural networks, deep ensembles, Monte Carlo (MC) dropout, and stochastic weight averaging-Gaussian (SWAG).
• Evaluated model performance based on empirical accuracy, reliability of confidence estimates, and practical applicability.
• Incorporated adversarial training and leveraged temporal/spatial correlations in air quality data.

Main Results:

• Proposed models demonstrated superior uncertainty quantification compared to previous methods in air quality forecasting.
• Bayesian neural networks provided more reliable uncertainty estimates but posed implementation challenges.
• Scalable methods (deep ensemble, MC dropout, SWAG) performed well with varying tradeoffs and performance metrics.

Conclusions:

• Uncertainty estimation significantly enhances the practical impact of data-driven air quality forecasts.
• Probabilistic models are more suitable for informed decision-making in air quality management.
• The choice of method involves balancing reliability, scalability, and implementation complexity.