Predicting biomass composition and operating conditions in fluidized bed biomass gasifiers: An automated machine learning approach combined with cooperative game theory

Abstract

A fluidized bed system for biomass gasification is considered a useful implementation with advantages of good heat and mass transfer and uniform heat across the bed that can reduce local hot spot which produces heavy hydrocarbons including tar. Such a system has been studied for decades for product conversion and selectivity optimization with empirical correlations, mathematical modelling, and machine learning. Although machine learning has been recently adopted for predicting biomass composition and operating conditions, the number of algorithms chosen from previous research is limited due to various types of models and complexity. An automated machine learning (AutoML) was adopted here to select the best machine learning algorithm amongst various types of models including tree ensembles and neural networks. Using AutoML, operating conditions and lignocellulosic compositions were predicted with output features from the system, including syngas composition, LHV, char yield, and tar yield. Generally, CatBoost (gradient boosting on decision trees) algorithm showed a good match with experimental data results/test data with high R2 (0.689) and low RMSE (0.220). Combined cooperative game theory (Shapely additive explanation, SHAP) was also applied to develop an interpretable model. AutoML combined with the SHAP algorithm for explainable machine learning was first tried in the field of fluidized bed systems to find a suitable machine learning model for each feature with hyperparameters optimization and expected to help in interpreting the results with limited results that require exhaustive experimentation. The results can be widely adapted in various scenarios, such as monitoring the process as a soft sensor.