Author: Bin CHEN (Xian Jiaotong University) - To predict H2, CO, CH4, and CO2 yields and low heat value (LHV), five machine learning (ML) models based on the experimental data of supercritical water gasification (SCWG) of straw were constructed. The multi-layer perceptron (MLP) is most suitable for H2 yield prediction, with high prediction accuracy (R2>0.925). The model that performed best in predicting CO yield is the random forest (RF) model. In predicting CH4 yield and CO2 yield, the AdaBoost model performed very well on both the training and testing sets. When predicting LHV, the prediction performance of RF is better than other models (R2＞0.963). Using the learning curve to judge the fitting state of the model, five different ML models have excellent performances in predicting LHV and can predict the experimental results more accurately. The influence degree of features on target variables was also compared. The influence order of operation conditions in SCWG of straw is that temperature (T) > concentration (c) > time (t). Finally, the generalized performance of the model is verified by the experimental data of bark SCWG, which proves that the ML model can be effectively extended to the prediction of hydrogen yield of bark or other raw materials. The present research shows that the ML model is a practical tool to predict the SCWG process, which helps to improve the efficiency of experiments and production.
