A hybrid data-driven machine learning framework for predicting the performance of coal and biomass gasification processes

Abstract

Gasification technology can effectively improve the utilization efficiency of coal and biomass resources. However, conventional experimental methods are costly, time-consuming, and labor-intensive to optimize the system performance of the different coal or biomass gasification process. Therefore, this study developed a hybrid data-driven machine learning framework for predicting the performance of coal and biomass gasification processes. To select the best machine learning model for the gasification process, the artificial neural network (ANN), decision tree, multiple linear regression, and support vector machine models are established with the hybrid database and assessed by seven regression evaluation indicators. The results indicate ANN model has the best prediction performance because it has the highest coefficient of determination (0.9242). To improve the prediction accuracy of the ANN model, the number of its hidden layers and neurons is first investigated and optimized. The results indicate that the preferred network structure of the ANN model is a double hidden layer neural network with 24 neurons. A genetic algorithm is then employed to improve the prediction performance of the optimized ANN model, which can further reduce the error of the ANN model. Finally, the genetic algorithm-optimized ANN model is applied to analyze the actual coal and biomass gasification processes. Results show that anthracite coal mixed with pine sawdust has the most significant impact on the gas yield of the gasification process, and bituminous coal mixed with rice husk has the most significant impact on the lower heating value of gasification process. Although the model has good predictive performance, it can continue to be improved by considering different equivalence or gasification ratios.