Development and Evaluation of Deep Learning Models for Predicting Instantaneous Mass Flow Rates of Biomass Fast Pyrolysis in Bubbling Fluidized Beds

Computational fluid dynamics (CFD) has evolved into a vital tool for advancing bubbling fluidized-bed reactors for biomass fast pyrolysis. However, due to the enormous computational burden of CFD simulations, optimizing working parameters over a broad range or simulating large/industrial units is still extremely time-consuming. Because deep learning (DL) is a promising method to attain both precision and speed, two new DL models, which added an attention mechanism or a convolutional neuron network (CNN) layer in the basic long short-term memory (LSTM) model, were established to predict instantaneous mass flow rates of major species for biomass fast pyrolysis in a bubbling fluidized bed. Historical mass flow rates from a multifluid model (MFM) simulation were considered as the time series of data for the model training process. Influencing factors, including sequence length, learning rate, convolutional kernel and stride sizes in the CNN layer, and number of neurons and layers in LSTM module, were examined to improve forecasting ability. The results demonstrated that the hybrid model including both CNN and LSTM outperforms other models in predicting instantaneous mass flow rates of biomass fast pyrolysis in bubbling fluidized beds.