Forecasting PM 2.5 concentration based on integrating of CEEMDAN decomposition method with SVM and LSTM

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Urbanization and consumption increase air pollution, particularly particulate matter (PM 2.5).
• Accurate PM 2.5 forecasting is essential for mitigating health risks and promoting sustainable development.
• Existing forecasting methods struggle to capture the complex dynamics of PM 2.5 concentrations.

Purpose of the Study:

• To develop and evaluate a novel hybrid model for accurate PM 2.5 concentration forecasting.
• To combine Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) for enhanced prediction.
• To assess the model's performance against traditional methods using various statistical metrics.

Main Methods:

• Utilized CEEMDAN to decompose PM 2.5 time series into Intrinsic Mode Functions (IMFs).
• Applied SVM and LSTM regression models to forecast individual IMF components.
• Employed the Naive Evolution algorithm for optimizing model parameters and combining forecasts.
• Trained and validated the model using daily PM 2.5 data from Kaohsiung, Taiwan (2019-2021).

Main Results:

• The proposed CEEMDAN-SVM-LSTM model demonstrated superior performance in 1-day ahead PM 2.5 forecasting.
• Achieved low Mean Absolute Error (MAE) of 1.858, Mean Square Error (MSE) of 7.2449, and Root Mean Square Error (RMSE) of 2.6682.
• Attained a high coefficient of determination (R²) of 0.9169, indicating excellent model fit.
• The model also showed the best performance for 3- and 7-day ahead PM 2.5 predictions.

Conclusions:

• The hybrid CEEMDAN-SVM-LSTM approach offers a robust and accurate method for PM 2.5 forecasting.
• This advanced forecasting capability can support environmental monitoring and sustainable urban planning.
• The model's effectiveness in predicting longer-term PM 2.5 trends warrants further investigation.