The employment of domain adaptation strategy for improving the applicability of neural network-based coke quality prediction for smart cokemaking process

Precise coke quality prediction is essential for coke production process optimization to achieve the reduction in energy consumption and CO₂ emissions, thus moving toward carbon neutrality in the coking industry. However, the complexity of coal molecular structures and the chemical reactions in the cokemaking process pose significant challenges to the applicability of existing coke quality prediction models. Based on the recently widely employed Artificial Neural Network (ANN) method, this study is the first to introduce domain adaptation strategies for improving the applicability of ANN in predicting coke quality including Coke Strength after Reaction (CSR) and Coke Reactivity Index (CRI). 649 Chinese coal samples with properties including M_ad, A_d, V_daf, S_t,d, G, X and Y along with coke quality were collected. They were initially categorized into source and target domains characterized by different distributions. Subsequently, two scenarios were independently evaluated based on coal samples in the target domain, which either lacked any information or contained partial information of actual coke quality. The results suggested that the proposed approach can significantly enhance the predictive performance for coal samples across various distributions. Moreover, a comprehensive investigation was also conducted to determine key factors influencing the effectiveness of coke quality prediction with this approach.