PM2.5 concentration prediction using weighted CEEMDAN and improved LSTM neural network

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Accurate prediction of fine particulate matter (PM2.5) concentration is critical for pollution management and public health.
• The inherent non-stationarity and nonlinearity of PM2.5 data present significant challenges for precise forecasting.
• Existing prediction models often struggle to capture the complex dynamics of atmospheric pollutant concentrations.

Purpose of the Study:

• To develop an advanced PM2.5 concentration prediction method that addresses data non-stationarity and nonlinearity.
• To enhance the accuracy and reliability of PM2.5 forecasting models.
• To improve the efficiency and global optimization capabilities of neural network hyperparameter tuning.

Main Methods:

• A novel Weighted Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (WCEEMDAN) method was employed to decompose PM2.5 sequences into weighted sub-layers.
• An Improved Long Short-Term Memory (ILSTM) neural network was utilized for sequence prediction.
• Adaptive Mutation Particle Swarm Optimization (AMPSO) was developed to optimize ILSTM hyperparameters, enhancing convergence speed and accuracy.

Main Results:

• The proposed WCEEMDAN-ILSTM model demonstrated superior performance in PM2.5 concentration prediction compared to other benchmark methods.
• Experimental results validated the model's effectiveness across three distinct PM2.5 datasets.
• The AMPSO algorithm significantly improved the optimization convergence and accuracy of the ILSTM network.

Conclusions:

• The WCEEMDAN-ILSTM model offers a robust and accurate approach for PM2.5 concentration prediction.
• The integration of WCEEMDAN and ILSTM effectively captures the complex characteristics of PM2.5 data.
• The study highlights the potential of advanced AI techniques for environmental monitoring and pollution control.