A Variational Bayesian Deep Network with Data Self-Screening Layer for Massive Time-Series Data Forecasting.

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Data-driven modeling is increasingly popular for forecasting, but big data can introduce noise and reduce accuracy.
• Traditional methods struggle with the complexity and inconsistency of large time-series datasets.

Purpose of the Study:

• To develop an advanced deep learning model for improved air quality forecasting.
• To address the challenges of noise, redundancy, and inconsistency in big time-series data for predictive modeling.

Main Methods:

• A novel deep network incorporating a data self-screening layer (DSSL) using maximal information distance coefficient (MIDC) to filter data.
• Utilizing a variational Bayesian gated recurrent unit (VBGRU) to enhance model robustness and noise resistance.

Main Results:

• The proposed model demonstrated superior accuracy in 24-hour PM2.5 concentration forecasting compared to existing models.
• The DSSL effectively filtered redundant and noisy data, improving overall model performance.

Conclusions:

• The developed deep network offers a superior approach to air quality forecasting by intelligently managing big data.
• The combination of data screening and robust recurrent units significantly enhances predictive accuracy and reliability.