Transient fluidization of biochar and sand mixtures for hydrogen production by steam reforming: a WL-CNN-LSTM deep learning study

Abstract Biochar is a promising feedstock for hydrogen production via steam reforming. However, its irregular morphology and low density lead to heterogeneous flow patterns in fluidized beds. This study presents a hybrid Wavelet-Enhanced Convolutional Neural Network–Long Short-Term Memory (WL-CNN-LSTM) framework for flow regime identification and pressure fluctuation prediction in fluidized beds of biochar-sand mixtures. Utilizing a Daubechies 4 (db4) wavelet for feature extraction, the model achieved 98.3 % accuracy in flow regime identification, representing a 4 % improvement over a standard CNN-LSTM model. For pressure prediction, the WL-CNN-LSTM framework reduced the computation time by over 52 % compared to a standalone LSTM in single-step forecasting, while its accuracy in multi-step prediction surpassed that of the CNN-LSTM by 33.5 %. The framework demonstrates a superior capability to capture transient hydrodynamics, showing great potential for the real-time monitoring and control of fluidized bed reactors to optimize hydrogen production.