Coal Calorific Value Prediction via Multi-View Transformer.

Area of Science:

• Analytical Chemistry
• Machine Learning
• Materials Science

Background:

• Accurate coal calorific value measurement is vital for efficient power generation.
• Conventional models struggle with large, heterogeneous coal datasets.
• Need for advanced analytical techniques for real-time coal quality assessment.

Purpose of the Study:

• To develop a novel deep learning framework, the Multi-View Transformer (MVFormer), for precise coal calorific value prediction.
• To integrate Near-Infrared Spectroscopy (NIRS) and X-ray Fluorescence (XRF) data for enhanced analysis.
• To improve upon existing models for large-scale, complex coal datasets.

Main Methods:

• Proposed a dual-pathway Transformer architecture (MVFormer) with Masked Autoencoder pre-training.
• Utilized fused NIRS and XRF data from over 20,000 coal samples.
• Implemented a multi-view fusion mechanism for improved generalization and feature representation.

Main Results:

• The MVFormer framework demonstrated significantly superior performance compared to Partial Least Squares (PLS) regression and Multilayer Perceptron (MLP) models.
• Achieved precise prediction of coal calorific value on large-scale datasets.
• Validated the framework's robustness and accuracy for industrial applications.

Conclusions:

• The MVFormer framework offers a robust and precise solution for real-time industrial coal quality analysis.
• Deep learning, particularly with fused spectroscopic data, enhances coal calorific value prediction accuracy.
• The study validates the effectiveness of advanced AI models for complex material characterization.