Deep Flexible Sequential (DFS) Model for Air Pollution Forecasting

Area of Science:

• Environmental Science
• Data Science
• Public Health

Background:

• Rapid urbanization and industrialization contribute to increasing air pollution in metropolitan areas.
• Air pollution, particularly PM2.5, poses significant risks to public health, necessitating accurate forecasting.
• Contextual factors like weather and traffic influence air quality, requiring adaptable models.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for forecasting PM2.5 concentrations.
• To provide timely air quality information to protect public health in urban environments.
• To create a flexible model applicable to diverse air pollution time-series datasets.

Main Methods:

• Designed and implemented a hybrid deep learning model named DFS (Deep Flexible Sequential).
• Integrated Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) architectures.
• Utilized flexible Dropout layers for enhanced generalization and adaptability to varying data window sizes.
• Trained and tested the model on real-world hourly PM2.5 data from Istanbul, Turkey (2014-2018).

Main Results:

• The DFS model successfully forecasts PM2.5 air pollution levels 4, 12, and 24 hours ahead.
• The model demonstrates generalization capabilities through its flexible architecture.
• The hybrid LSTM-CNN approach effectively captures complex temporal dependencies in air quality data.

Conclusions:

• The DFS model offers a robust and flexible solution for short-term air quality forecasting.
• Accurate PM2.5 prediction is crucial for public health advisories in urban settings.
• The model's adaptability allows for application to other air pollution time-series forecasting challenges with minor parameter adjustments.