Abstract: To address the significant errors associated with conventional neural network models in predicting the fusibility of coal ash with complex composition, a novel weighted coupled neural network model (WCN) is proposed. It is particularly characterized by the modification of the traditional feedforward neural network based on the competitive reaction mechanisms of the coal ash Si-Al-Ca-Fe systems. Specifically, the silica-to-alumina ratio (S/A) and calcium-to-iron ratio (C/F) of coal ash are incorporated as feature items into a weighted network, whose outputs are then coupled with those of the component network, enabling the neural network to have flexible and adjustable prediction weights, thereby enhancing the model's adaptability and prediction accuracy. Comparison of the prediction results of coal ash softening temperature by different models demonstrates that the maximum errors of the WCN model in the prediction of the test set is below 60 ℃, outperforming the reproducibility requirement (80 ℃) for coal ash fusion characteristics in the Chinese National Standard(GB/T 219—2008). Furthermore, compared to the conventional eXtreme Gradient Boosting(XGBOOST) and back propagation neural network(BPNN) models, the WCN model improves prediction accuracy for complex high-alkali coal(e.g., Zhundong coal) ash by 32.8% and 83%, respectively. The study shows that the WCN model achieves significant improvements in both coal ash adaptability and prediction accuracy, demonstrating considerable application value.

Key words: coal ash, fusibility, softening temperature, neural network, model