Particulate matter emission area identification based on phenomenological atmospheric dispersion and deep learn...

Area of Science:

• Environmental Science
• Atmospheric Science
• Data Science

Background:

• Industrial and port areas pose significant environmental and public health risks due to multiple particulate matter (PM) emission sources.
• Existing real-time monitoring and predictive models are insufficient for precise PM emission source detection.

Purpose of the Study:

• To develop an integrated framework for precise, real-time localization of PM emission sources in flat terrain.
• To enhance environmental regulation, industrial accountability, and public health protection through improved PM source identification.

Main Methods:

• Developed an integrated framework combining Artificial Neural Networks (ANNs) and Computational Fluid Dynamics (CFD).
• Validated the CFD model using experimental data and Monin-Obukhov similarity theory for accurate atmospheric profiles and PM transport.
• Trained Long Short-Term Memory (LSTM) and 1D Convolutional Neural Network (CNN1D) models on a CFD-generated dataset of 243 simulation runs.

Main Results:

• Both LSTM and CNN1D deep learning models achieved high accuracy, with F1-scores exceeding 0.95 for PM emission source classification.
• The framework demonstrated reliable emission source localization capabilities.
• Hyperparameter optimization times differed: LSTM (4h 15min) and CNN1D (4h 43min).

Conclusions:

• The integrated CFD-ANN framework offers a reliable solution for real-time PM emission source localization.
• This approach shows significant promise for improving environmental management and public health strategies in industrial and port environments.
• Represents a breakthrough in addressing real-time PM source identification challenges.