Dynamic Modeling for Temperature Prediction in a Fluidized Bed Biomass Gasification Process

Abstract

A long short-term memory (LSTM) based artificial neural network (ANN) model is developed to predict the temperatures of the fluidized bed and freeboard as well as of the outlet gas temperature during the fluidized bed biomass gasification process. The various available equilibrium models and the kinetic models to simulate this process are often found impractical at a large scale due to the assumptions on which these models are framed. Instead of solving complex kinetic and equilibrium models, this study explores the potential of utilizing an LSTM model for the dynamic modeling of a fluidized bed biomass gasifier. The influence of various hyperparameters on LSTM model performance are also studied. Among many tested activation functions, tanh has the lowest mean squared error overall. Adagrad and SGD are identified as the most suitable optimizers for our LSTM model. Single and two layered LSTM models are compared. The LSTM model is validated with the spatio-temporally resolved experimental data of the pilot plant fluidized bed gasifier. The LSTM model is further examined for temperature predictions at three future points (1 min, 3 min, and 5 min ahead). Overall, the LSTM model offers a sufficiently accurate alternative for temperature prediction in fluidized bed biomass gasifier.