PM2.5 concentration modeling and prediction by using temperature-based deep belief network.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Particulate Matter (PM2.5) significantly impacts air quality globally.
• Accurate PM2.5 prediction remains challenging due to complex formation factors.

Purpose of the Study:

• To propose a novel deep learning model, Temperature-based Deep Belief Networks (TDBN), for predicting daily PM2.5 concentrations.
• To evaluate the TDBN model's performance against other models using real-world data.

Main Methods:

• Utilized Partial Least Square (PLS) to select auxiliary input variables for the TDBN model.
• Developed TDBN using temperature-based Restricted Boltzmann Machines (RBM), incorporating temperature as a physical parameter.
• Optimized TDBN structural parameters (hidden layers, neurons, temperature value) by minimizing training errors.

Main Results:

• The TDBN model demonstrated superior performance in PM2.5 prediction compared to similar models.
• TDBN achieved better results in Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²).

Conclusions:

• The proposed TDBN model offers a promising approach for enhanced PM2.5 air quality prediction.
• Incorporating temperature as a physical parameter in deep learning models can improve prediction accuracy.