Métodos de inteligência artificial aplicados em dados de biomassa para a caracterização dos diferentes tipos de pirólise

Abstract

This study addresses a biomass data classification problem. One of the objectives is to identify the most relevant variables for classifying the pyrolysis type of biomass. Also, to evaluate if the pyrolysis type classes are sufficient to characterize this chemical process. The Random Forest algorithm was applied to identify which variables are relevant in the pyrolysis-type classification process, obtaining an accuracy of around 97%. It was identified that the most important variables are: Average residence time in the reactor for the gas and carrier, Percentage of ash-free dry basis carbon in the feedstock, Average particle size in the reactor, and Percentage of ash-free dry basis hydrogen in the feedstock. The clustering methods were used with the most relevant variables: k-means, pam, clear, diana, fanny, hierarchical, sound, sota, and model. To evaluate the clustering methods, internal validation measures with Dunn’s index metrics and silhouette were used. The validation measures indicated the hierarchical clustering and k-means with better results for several groups greater than the number of existing pyrolysis classes. Thus, the dataset can be divided into more pyrolysis-type groups, considering only the available classes are too limited to characterize the pyrolysis type since the unsupervised classification algorithms indicate the number of clusters as greater than or equal to five.