Deep-learning architecture for PM_2.5 concentration prediction: A review.

Area of Science:

• Environmental Science
• Computer Science
• Data Science

Background:

• Accurate prediction of fine particulate matter (PM2.5) is critical for assessing air pollution and public health risks.
• Deep learning (DL) models are increasingly used for PM2.5 forecasting, but a standardized evaluation framework is lacking.

Purpose of the Study:

• To systematically review deep learning-based hybrid models for PM2.5 forecasting.
• To categorize and compare the performance, complexity, and interpretability of various DL models.
• To introduce a novel evaluation framework (DMES) for standardizing DL-based PM2.5 prediction models.

Main Methods:

• Systematic literature review following PRISMA guidelines.
• Categorization of PM2.5 DL methodologies into seven types (four DL-based, three hybrid).
• Comparative analysis of model complexity, effectiveness, innovation, and interpretability.

Main Results:

• Established DL architectures are efficient but often lack innovation and interpretability.
• Hybrid models with traditional approaches offer interpretability but sacrifice accuracy/speed.
• Novel hybrid DL models show innovation but face interpretability challenges.

Conclusions:

• A unified and standardized evaluation framework is needed for DL-based PM2.5 prediction.
• The proposed Dataset-Method-Experiment Standard (DMES) framework aims to standardize model evaluation.
• This work provides a foundation for improving the quality and standardization of DL model usage in PM2.5 prediction research.