Prediction of instantaneous yield of bio-oil in fluidized biomass pyrolysis using long short-term memory network based on computational fluid dynamics data

Abstract

Computational fluid dynamics (CFD) is an effective tool to investigate biomass fast pyrolysis in fluidized bed reactor for bio-oil production, while it requires huge computational time when optimizing operating conditions or simulating large/industrial units. Machine learning (ML) is a promising approach to achieving both accuracy and efficiency. In this work, a reduced-order model including long short-term memory (LSTM) layer, pooling layer, and fully connected layer was established to predict future mass flow rates by training the historical CFD data. Unsteady mass flow rates, which are normally used to determine product yields, were treated as data of time series in ML. Influencing factors, such as sequence length, number of neurons, learning rate, subsequences order (shuffle or not), number of LSTM layers, and ratio of testing set, were evaluated to obtain their optimal values. The developed LSTM model framework and training process showed good applicability for the dataset of different species and temperatures. Product yields predicted by the derived LSTM were in good agreement with those obtained by CFD, while nearly 30% computational effort was saved. Thus, it is clearly seen that the well-predicted fluctuating characteristics and final product yields are helpful to improve accuracy of process simulation for digitalizing key reactors and building smart factories.