A Spatial-Temporal Causal Convolution Network Framework for Accurate and Fine-Grained PM2.5 Concentration Prediction | JoVE Visualize

Area of Science:

Environmental Science
Data Science
Atmospheric Chemistry

Background:

Accurate prediction of PM2.5 concentration is crucial for air quality control and public health.
Existing methods like RNNs and GCNs struggle to integrate complex spatial-temporal and meteorological factors.
Dynamic relationships between scattered monitoring stations present a challenge for traditional models.

Purpose of the Study:

To propose a novel spatial-temporal causal convolution network framework, ST-CCN-PM2.5, for fine-grained PM2.5 prediction.
To effectively integrate multi-source air pollutant and meteorological data using attention mechanisms.
To enhance the prediction accuracy by capturing time-dependent features and optimizing model hyperparameters.

Main Methods:

Developed the ST-CCN-PM2.5 framework incorporating spatial attention for multi-source pollutant and meteorological factors.
Utilized stacked dilated convolution and time attention to extract time-dependent features within causal convolution networks.
Employed Bayesian optimization for hyperparameter tuning and validated the model using Haikou air monitoring station data against various baselines.

Main Results:

ST-CCN-PM2.5 demonstrated significant performance improvements, with average reductions in RMSE, MAE, and R² of 27.05%, 10.38%, and 3.56% for single-station predictions, respectively.
Across all stations, ST-CCN-PM2.5 achieved superior performance in win-tie-loss experiments, excelling in RMSE, MAE, and R² metrics.
Shapley analysis identified wind speed as the most influential factor, with CO and temperature also showing moderate significance in PM2.5 prediction.

Conclusions:

The ST-CCN-PM2.5 framework offers a promising advancement for accurate and fine-grained PM2.5 concentration prediction.
The model's ability to integrate diverse data sources and capture complex spatial-temporal dynamics is key to its effectiveness.
Findings highlight the importance of meteorological factors, particularly wind speed, in predicting PM2.5 levels.