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

The combined operation of multiple particulate matter (PM) emission sources in industrial and port areas creates major environmental threats and serious public health risks. Current methods of monitoring and predictive models lack sufficient capability to detect PM emission sources in real time. This study developed an integrated framework that uses Artificial Neural Networks (ANNs) and Computational Fluid Dynamics (CFD) to precisely locate PM emission sources in flat terrain. The CFD model was validated through experimental data analysis and the Monin-Obukhov similarity theory to precisely represent the particulate matter transport and atmospheric profiles. We created a simulation dataset containing 243 runs that tested different wind speed and direction combinations with variations in emission height and emission interval. The dataset served as training material for two deep learning models which used Long Short-Term Memory (LSTM) and a one-dimensional Convolutional Neural Network (CNN1D) to perform PM emission location classification. Both models achieved high accuracy levels with F1-scores above 0.95. The time needed to optimize hyperparameters proved the difference between models because LSTM required 4 h and 15 min and CNN1D needed 4 h and 43 min. This study proves that using CFD-generated data with ANN models allows reliable emission source localization which shows promise for environmental regulation, industrial accountability, and public health protection. The proposed framework represents a major breakthrough in real-time PM source localization in industrial and port environments.