The novel adaptive graph neural network-based coke quality prediction for coal samples with missing properties in sustainable smart cokemaking applications

• A novel graph expression-based method was proposed for handling missing coal properties. • Introducing a novel residual graph attention network approach for coke quality prediction. • The model demonstrates excellent performance on Chinese and Australian coal samples. • The influencing factors in the graph-based coke quality prediction were explored. Although numerous coke quality prediction models have been proposed in the past decades, they struggle to provide predictions for coal samples with varying numbers of properties from various sources. Predictions cannot be made when input coal samples containing missing values fail to align with the trained model’s expected input properties. A recent study showed that image-based expression of coal properties can achieve superior performance compared to traditional numerical coal properties, but it still faces this challenge. To address this, this study is the first to introduce a novel method that converts numerical coal properties into graph expressions, with the number of nodes and edges dynamically adjusting based on input properties. A residual graph attention network was developed to process these graphs for Coke Reactivity Index (CRI) and Coke Strength after Reaction (CSR) prediction. The model was trained and tested on 808 Chinese coal samples and further validated on 20 Australian and 38 Russian coal samples. The experimental results indicated that it can effectively handle coal samples with missing values and outperformed other regression approaches and models integrated with various missing value imputation strategies, achieving Mean Absolute Error (MAE) values of 2.08 and 2.65 for predicting CRI and CSR in Chinese coal samples, while also demonstrating excellent applicability to Australian coal samples. However, its performance was limited on Russian coal samples due to significant distributional differences. Furthermore, it can provide insights into interactions among coal properties during prediction, improving model interpretability in correlating CRI and CSR. A comprehensive investigation was further carried out to examine crucial factors affecting model performance.