Prediction of PM_2.5 concentration based on a CNN-LSTM neural network algorithm.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Fine particulate matter (PM2.5) poses significant risks to human health and survival.
• Accurate PM2.5 concentration prediction is crucial for source tracing and implementing public health measures.

Purpose of the Study:

• To predict and analyze PM2.5 concentration using an integrated deep learning Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model.
• To address the complexity and nonlinear characteristics of PM2.5 time series data.

Main Methods:

• Collected PM2.5 and meteorological data (temperature, wind speed, air pressure) for Qingdao in 2020.
• Employed a CNN-LSTM deep learning model to extract spatial features (CNN) and temporal dependencies (LSTM).
• Conducted comparative experiments against CNN and LSTM models for performance evaluation.

Main Results:

• The CNN-LSTM model achieved excellent PM2.5 prediction performance with R² of 0.91 and RMSE of 8.216 µg/m³.
• Demonstrated superior accuracy and generalizability compared to standalone CNN (R²=0.85, RMSE=11.356) and LSTM (R²=0.83, RMSE=14.367) models.
• Identified significant effects of meteorological factors (temperature, pressure, wind speed) on ground-level PM2.5 concentrations.

Conclusions:

• The integrated CNN-LSTM deep learning approach provides superior PM2.5 prediction accuracy.
• Revealed significant associations between PM2.5 concentrations and key meteorological factors.
• Findings support improved atmospheric environment quality management and public health protection strategies.