Predicting gas production by supercritical water gasification of coal using machine learning

Supercritical water gasification of coal is a potential clean conversion technology. Applying machine learning (ML) methods can reduce costs and avoid the distortion of mechanism models, which has attracted increasing attention. This paper collected 208 experimental samples, including a total of 3536 data points used as a data set to investigate six independent ML models. A 5-fold cross-validation method combined with grid search was used to obtain the optimal hyperparameter combination. The overall performance ranking of the six developed models is GBR > RF > SVR > DT > ANN > ABR. The features were analyzed using the interpretable model with SHAP values, which showed that the contribution of operating conditions to the gas yield reached 88.55 %, and coal properties to gas yield was only 11.45 %. The top three models with the best prediction performance of each gas were weighted and combined to establish a hybrid model. The performance of the hybrid model on the test set is improved compared with the original GBR model. The carbon gasification efficiency of 17 supplementary experimental samples outside the dataset was predicted using the hybrid model. The MRE of 17.92 % and the R2 of 0.920 were obtained, showing a solid generalization ability.