Machine learning methods for modelling the gasification and pyrolysis of biomass and waste

• Published in: Renewable & sustainable energy reviews 2022-03, Vol.155, p.111902, Article 111902
• Main Authors: Ascher, Simon, Watson, Ian, You, Siming
• Format: Article
• Language: English
• Subjects: Biomass; Gasification; Machine learning; Pyrolysis; Waste
• ISSN: 1364-0321; 1879-0690
• DOI: 10.1016/j.rser.2021.111902

Over the past two decades, the use of machine learning (ML) methods to model biomass and waste gasification/pyrolysis has increased rapidly. Only 70 papers were published in the 2000s compared to a total of 549 publications in the 2010s. However, the approaches and findings have yet to be systematically reviewed. In this work, the machine learning methods most commonly employed for modelling gasification and pyrolysis processes are discussed with reference to their applications, merits, and limitations. Whilst coefficients of determination (R2) can be difficult to compare directly, due to some studies having greatly different approaches and aims, most studies consistently achieved a high prediction accuracy with R2 > 0.90. Artificial neural networks have been most widely used due to their potential to learn highly non-linear input-output relationships. However, a variety of methods (e.g. regression methods, tree-based methods, and support vector machines) are appropriate depending on the application, data availability, model speed, etc. It is concluded that ML has great potential for the development of models with greater accuracy. Some advantages of machine learning models over existing models are their ability to incorporate relevant non-numerical parameters and the power to generate a multitude of solutions for a wide range of input parameters. More emphasis should be placed on model interpretability in order to better understand the processes being studied. •Machine learning models can accurately model gasification and pyrolysis.•Artificial neural networks (ANN) have most frequently been used.•Black box nature and lack of interpretability are drawbacks of ANN.•Particle swarm optimisation used for selecting network structure and hyperparameters.•Data availability for model development remains a key challenge.