Machine learning prediction and analysis of commercial wood fuel blends used in a typical biomass power station. (15th May 2022)

• Record Type: Journal Article
• Authors: Morris, Jonathan D.; Daood, Syed Sheraz; Nimmo, William
• Is Part Of: Fuel. Volume 316(2022)
• Publication Date: 2022-05-15
• Subjects: Biomass -- Waste wood -- Ash -- Random forest; Fuel -- Periodicals; Coal -- Periodicals; Coal; Fuel; Periodicals; 662.6
• DOI: 10.1016/j.fuel.2022.123364
• Languages: English
• ISSNs: 0016-2361

Abstract

Graphical abstract: Highlights: 5 year blended virgin wood/recycled wood data set from biomass power station. Analysis of Ash, K, Na, Cl, Pb, Zn; prediction with random forest (RF) regression. Ash and Cl quarterly levels exceed literature average >50% of the time. RF regression created poor predictive models; R 2 generally <0.6 (except ash 0.787). Suggestions regarding application of ML methods for ash & trace element prediction. Abstract: Biomass can be utilised as a near carbon neutral fuel for power generation. Biomass may come in a variety of forms, such as woods or agricultural wastes, with highly variable compositions. It is well known that biomass ash content and composition can introduce severe operational challenges to power stations, such as slagging and corrosion. Machine learning approaches have been successfully applied in a variety of contexts to generate predictive models and identify relationships in large data sets. However, such approaches have not previously been applied to biomass trace element or ash content data, in part due to limited large data set availability. In this work, 5 years of fuel blend analysis data (3500+ data sets) was analysed from a 35MWe biomass power station burning a 60/40% blend of virgin wood and recycled (waste) wood. Variation to ash content and key trace elements (K, Na, Pb, Zn, Cl) was analysed and compared versus a literature average benchmark. Identification of underlying relationships between these key components and others was attempted with principal component analysis and random forest regression machine learning. Ash and chlorine exceeded the literature average benchmark over many time periods and would have a negative impact on boiler operation. Potassium and sodium were only above literature average levels intermittently. No significant underlying relationships for the key components could be identified with principal component analysis or random forest regression, nor could an accurate predictive model be created, though some minor trends were noted. This is likely due to the high degree of heterogeneity seen in the fuel data, as it is a blend of virgin and recycled wood. It is suggested that future studies applying machine learning methods in this context either use singular fuel data sets, or that additional information is recorded within the blended data set when analysis fuel composition (e.g. fuel sources, suppliers, blend ratio). Suggestions were also made regarding improvements to waste wood sampling approaches.