Dust Concentration Forecasting Method for Intermittent Processing of Powder and Granular Materials

Area of Science:

• Environmental Science
• Data Science
• Engineering

Background:

• Dust concentration in industrial settings exhibits abrupt changes and complex sensor interdependencies.
• Existing forecasting models struggle with global dependencies and local trend characterization.

Purpose of the Study:

• To develop an advanced dust concentration forecasting model using iTransformer.
• To enhance the modeling of transient variations and peak fluctuations in dust levels.

Main Methods:

• Proposed an iTransformer-based model with a dual-stage feed-forward network and DLinear branch.
• Implemented variate-wise modeling for multi-source sensing signal coupling.
• Utilized an adaptive gated fusion mechanism for dynamic branch contribution.

Main Results:

• Achieved superior performance on a high-frequency multivariate PM2.5 dataset.
• Demonstrated significant improvements in forecasting accuracy with low MSE, MAE, RMSE, and MAPE.
• The model outperformed baseline methods in overall forecasting performance.

Conclusions:

• The proposed model offers improved accuracy for sensor-driven particulate concentration forecasting.
• Provides a methodological reference for early warning systems in industrial environments.
• Further validation with field data is recommended for practical deployment.