Hourly PM_2.5 concentration prediction for dry bulk port clusters considering spatiotemporal correlation: A novel deep...

Area of Science:

• Environmental Science and Engineering
• Artificial Intelligence in Environmental Monitoring
• Air Quality Prediction Modeling

Background:

• Accurate prediction of PM_2.5 concentrations in port environments is vital for public health and effective air pollution control.
• Port clusters present unique challenges due to complex meteorological conditions, inter-port PM_2.5 correlations, and urban background pollutant influences.
• Existing models often struggle to capture the spatiotemporal dynamics inherent in port cluster air quality.

Purpose of the Study:

• To develop a novel blending ensemble deep learning model for accurate PM_2.5 concentration prediction in port clusters.
• To address the challenges posed by spatiotemporal correlations and external pollutant influences in port environments.
• To provide reliable decision support for air quality management in dry bulk port clusters.

Main Methods:

• Developed a blending ensemble deep learning model integrating Graph Convolutional Networks (GCN), Long Short-Term Memory (LSTM) networks, and Residual Neural Networks (ResNet).
• Utilized GCN for spatial correlation, LSTM for temporal dependencies, and ResNet for urban pollutant effects, with a CNN meta-model for final prediction.
• Validated the model using 18 ports in Nanjing, comparing its performance against six state-of-the-art models.

Main Results:

• The proposed GCN-LSTM-ResNet model demonstrated superior prediction accuracy, significantly outperforming existing methods.
• Achieved reductions in Mean Absolute Error (MAE) by 10.59%-20.00% and Root Mean Square Error (RMSE) by 13.22%-17.11%.
• Showcased improved Coefficient of Determination (R²) by 10%-35.38% and Accuracy (ACC) by 3.48%-7.08%, with notable performance in predicting high-concentration events.

Conclusions:

• The novel blending ensemble deep learning model offers a robust solution for PM_2.5 prediction in complex port cluster environments.
• GCN and LSTM components were identified as having the most significant impact on prediction performance, highlighting the importance of spatial and temporal factors.
• The model provides reliable predictions and valuable insights for developing effective air quality management strategies in port areas.