Chemical looping in packed beds is an inherently dynamic process that lacks a conventional steady state, therefore requiring dynamic optimization methods to maximize efficiency. Algebraic models can accurately represent such systems; however, the time required to solve them is high and they may not capture the behaviour fully, complicating rapid process study and optimization workflows. To address this challenge, surrogate models developed with deep learning methods are presented in this work. The models were trained using synthetic data obtained from an algebraic model of a chemical looping reactor developed in Aspen Custom Modeler. Only variables that can be observed or manipulated were included in the training data. Direct forecasting was conducted with an encoder–decoder Transformer and an LSTM was used as the baseline. The surrogate model formulation is explicitly conditioned on future controls, can take into account state information over long distances, is tractable for forecasts of multiple thousands of steps, and is capable of fast inference once trained. Depending on hardware, speedups of up to 30000×over algebraic simulations were observed, with the Transformer-based approach found to outperform LSTM both on accuracy and speed. The Transformer model achieved a mean squared error of 0.0051 on normalized data from the testing set, with most of the error resulting from relatively poor predictions of the slow-moving temperature profile. Overall, the accuracy of predictions is sufficient to evaluate reactor performance objectives. In addition to being potentially useful for future control-oriented applications with regard to chemical looping in packed beds, this approach does not use any chemical looping-specific information or structure, and may be transferable to other long-horizon dynamic process systems. • Surrogate models trained on synthetic data. • Control-conditioned masked modelling. • Up to 1 0 4 times faster prediction. • Test set MSE of 0.0051 and 0.0104 for Transformer and LSTM models, respectively. • LSTM baseline showed a very limited use of long-range context.
Zurba et al. (Fri,) studied this question.