High-accuracy PM_2.5 prediction via mutual information filtering and Bayesian-Optimized Spatio-Temporal Convolutional...

Area of Science:

• Environmental Science
• Data Science
• Computer Science

Background:

• Air pollution, especially fine particulate matter (PM2.5), presents significant risks to public health and ecosystems.
• Accurate PM2.5 concentration prediction is vital for public health interventions and policy-making.
• Deep learning models often struggle with raw data, leading to feature redundancy, prolonged training, overfitting, and reduced prediction accuracy.

Purpose of the Study:

• To develop an advanced PM2.5 prediction framework that addresses challenges of feature redundancy and model optimization.
• To enhance the accuracy and efficiency of PM2.5 concentration forecasting.
• To improve the robustness and generalization capabilities of spatiotemporal prediction models.

Main Methods:

• A dynamic feature selection framework integrating Mutual Information (MI) and Adaptive Information Distance (AID) to prune redundant inputs.
• A Bayesian optimizer utilizing multimodal Gaussian distributions for efficient global hyperparameter search and improved model robustness.
• An Information Screening-Enhanced Spatiotemporal Convolutional Network (MIBO-STCN) combining causal convolution, adaptive receptive fields, and an information screening layer.

Main Results:

• The proposed framework adaptively prunes redundant features, enhancing the information utility of the input data.
• The Bayesian optimizer effectively explores the parameter space, leading to better hyperparameter selection and model stability.
• The MIBO-STCN model demonstrates synergistic optimization of spatiotemporal dependency modeling and redundancy reduction, significantly boosting prediction accuracy and accelerating convergence.
• Experimental results show the proposed approach outperforms state-of-the-art models in PM2.5 concentration forecasting across diverse scenarios.

Conclusions:

• The developed framework effectively reduces feature redundancy and optimizes model performance for PM2.5 prediction.
• The integration of dynamic feature selection, advanced Bayesian optimization, and an enhanced spatiotemporal network leads to superior forecasting accuracy and efficiency.
• The MIBO-STCN framework offers robust generalization capabilities, making it a valuable tool for environmental monitoring and public health policy.