Estimation of lignocellulosic biomass pyrolysis product yields using artificial neural networks

As the push towards more sustainable ways to produce energy and chemicals intensifies, efforts are needed to refine and optimize the systems that can give an answer to these needs. In the present work, the use of neural networks as modelling tools for lignocellulosic biomass pyrolysis main products yields estimation was evaluated. In order to achieve this, the most relevant compositional and reaction parameters for lignocellulosic biomass pyrolysis were reviewed and their effect over the main products yields was assessed. Based on relevant literature data, a database was set up, containing parameters and experimental results from 32 published studies for a total of 482 samples, including both fast and slow pyrolysis experiments performed on a heterogeneous collection of lignocellulosic biomasses. The parameters that in the database configured as best predictors for the solid, liquid and gaseous products were determined through preliminary tests and were then used to build reduced models, one for each of the main products, which use five parameters instead of the full set for the estimation of yields. The procedures included hyperparameter optimizations steps. The performances of these reduced models were compared to those of the ones obtained using the full set of parameters as inputs by using the root mean squared error (RMSE) as metric. For both the char and gas products, the best results were consistently achieved by the reduced versions of the network (RMSE 5.1 wt% ar and 5.6 wt% ar respectively), while for the liquid product the best result was given by the full network (RMSE 6.9 wt% ar) indicating substantial value in proper selection of the input features. In general, the char models were the best performing ones. Additional models for the liquid and gas product featuring char as additional input to the system were also devised and obtained better performance (RMSE 5.5 wt% ar and 4.9 wt% ar respectively) compared to the original ones. Models based on single studies were also included in order to showcase both the capabilities of the tool and the challenges that arise when trying to build a generalizable model of this kind. Overall, artificial neural networks were shown to be an interesting tool for the construction of setup-unspecific biomass pyrolysis product yield models. The obstacles standing currently in the way of a more accurate modelling of the system were highlighted, along with certain literature discrepancies, which hinder reliable quantitative comparison of experimental conditions and results among separate studies.