A new hybrid PM volatility forecasting model based on EMD and machine learning algorithms.

Area of Science:

• Environmental Science and Engineering
• Atmospheric Science
• Data Science and Machine Learning

Background:

• Air pollution, particularly PM2.5, poses significant risks to public health and life.
• Accurate prediction of PM2.5 volatility is crucial for effective air quality management.
• Existing machine learning models like LSTM and SVM often overlook crucial time-frequency information in volatility series.

Purpose of the Study:

• To develop a novel hybrid model for enhanced PM2.5 volatility prediction.
• To integrate time-frequency characteristics using Empirical Mode Decomposition (EMD).
• To incorporate residual and historical volatility via the Generalized AutoRegressive Conditional Heteroskedasticity (GARCH) model.

Main Methods:

• Empirical Mode Decomposition (EMD) for time-frequency characteristic extraction of PM2.5 volatility series.
• Generalized AutoRegressive Conditional Heteroskedasticity (GARCH) model to integrate residual and historical volatility.
• Hybridization of EMD and GARCH with machine learning algorithms (LSTM and SVM) for prediction.

Main Results:

• The hybrid-LSTM model demonstrated a reduction in Mean Absolute Deviation (MAE) from 0.00875 to 0.00718 compared to standalone LSTM.
• The hybrid-SVM model showed significant improvement in generalization ability, with the Index of Agreement (IA) increasing from 0.846707 to 0.96595.
• The proposed hybrid model outperformed benchmark models in prediction accuracy and stability across 54 North China cities.

Conclusions:

• The hybrid model effectively captures complex volatility patterns by leveraging time-frequency information.
• The integration of EMD and GARCH with machine learning offers superior PM2.5 volatility prediction capabilities.
• This hybrid system modeling approach is highly suitable for PM2.5 volatility analysis and air quality forecasting.