Causal Enhanced CoAtNet BiFPN Model for Coal Ash Prediction
Authors
Marcin Skowroński, Zuzanna Lorenc, Emilia LipińskaFiles
Abstract
Reliable ash-content prediction is needed for separation monitoring and product quality assessment in coal preparation plants, where moisture, particle size, and feed-source changes can alter sensing responses. This paper presents a causal- informed CoAtNet-BiFPN framework combining context-dependent sample weighting, convolution-attention feature extraction, and bidirectional multi-scale fusion. The numerical case study describes 12,480 synchronized image, near- infrared, and process-record samples spanning ash contents from 6.1% to 34.8%, with chronological training, validation, and test partitions. Context variables include moisture, dominant size fraction, belt speed, illumination, and source batch. The illustrative comparison reports a mean absolute error of 0.52 percentage points, a root mean squared error of 0.74 percentage points, a mean absolute percentage error of 3.18%, and a coefficient of determination of 0.972. Relative to the same backbone and fusion neck without context adjustment, the absolute error reductions are 0.09 and 0.11 percentage points, respectively; the latter corresponds to a 12.9% relative reduction. Under the reported held-out source condition, the proposed model's root mean squared error increases from 0.74 to 0.85 percentage points. These numerical comparisons illustrate the intended evaluation of predictive accuracy and contextual robustness, rather than establishing an identifiable causal effect or independently verified plant performance. The framework is intended for supervisory decision support; traceable assay records, implementation details, and external validation remain necessary before deployment claims can be substantiated.
