Investigation of factors affecting performance of a downdraft fixed bed gasifier using optimized MLP neural networks approach

Abstract

Biomass gasification will be a competitive renewable technology to meet the world's energy demand in the near future. However, extremely time-consuming and costly experimental investigations are still required to achieve the best gasification performance. Machine learning through artificial neural networks can be considered as a convenient and low-cost tool for predicting and optimizing gasification conditions. In this paper, MLP neural network was used to develop a predictive model for the fixed bed gasification properties. Several MLP models were developed to predict the composition of the produced gas and the lower heating value based on the physicochemical composition of the biomass and the reactor operating conditions. The results showed that the MLP architecture with the Levenberg-Marquardt algorithm by considering the range of 47 to 57 neurons in the hidden layer and the tansig activation function had reasonable accuracy in predicting the outputs. Performance of the MLP proposed model showed good agreement between the output and target values with a coefficient of determination of R2 > 0.952, root mean squared error, RMSE < 0.83, and relative root mean squared error, rRMSE < 6.5 %. In addition, according to the results obtained, the MLP indicated the highest precision to model the fixed bed gasification over the other methods. Finally, after the implementation of the model, the sensitivity analysis of input data on output data was performed. The modeling results of this study showed that the MLP model can effectively replace the costly experimental tests to study the gasification process.