Development and assessment of hybrid machine learning model of biomass pyrolysis process

Abstract

• An improved biomass pyrolysis process model using machine learning is proposed. • The model was developed as a hybrid Aspen Plus and XGBoost model. • XGBoost predicted pyrolysis products with R 2 ranging from 0.76 to 0.89. • XGBoost offered the highest accuracy compared to other modelling methods. • The output of pyrolysing 136 biomass species can be predicted from the model. Pyrolysis of lignocellulosic biomass is a potential pathway to produce hydrocarbon biofuels. To progress commercial development of the pyrolysis process, models are used to predict the output products, optimise the process, and investigate process feasibility. This work presents a hybrid model fusing Aspen Plus and a machine learning model (random forest, Artificial Neural Network, ANN, and eXtreme Gradient Boosting ensemble tree, XGBoost). The Aspen Plus part of the model aims to develop a comprehensive pyrolysis flowsheet while the machine learning aims to enhance the accuracy and agility of the model. Pyrolysis is frequently simulated in Aspen Plus software based on manual specification of product components. This approach simulates a case of a single feedstock and a corresponding set of operating parameters, making sensitivity analyses onerous as it requires several variables to be manually manipulated to simulate a different scenario. Random forest, ANN and XGBoost models were applied to improve the accuracy and agility of the model using empirical data of the gaseous, water, bio-oil yield and composition, and char yield collected from literature. XGBoost was found to offer lower RMSE than ANN and random forest when predicting output products. Hence, XGBoost was selected for the hybrid model. The coefficient of determination (R 2 ) results after optimising hyperparameters ranged from 0.76 to 0.89. Moreover, the XGBoost model developed in this study offered higher accuracy than the use of minimisation of Gibbs Free Energy and kinetic modelling methods. This model could reliably predict the production of upgraded bio-oil products from a broad range of biomass species, and calculate the energy requirements, electricity generated, and CO 2 emissions by specifying the biomass characteristics and pyrolysis temperature only. This approach serves as a key advance in developing reliable and flexible models to investigate economic feasibility and aid decision makers in developing biomass pyrolysis processes.

Topics & Concepts