Modeling of a sorption heat storage reactor using nonlinear autoregressive neural networks

Sorption thermal energy storage has the potential to store thermal energy over a long time with a higher energy density and less thermal losses compared to other technologies. In terms of modeling, sorption reactors are commonly described by physics-based models encompassing complex nonlinear phenomena occurring in the reactor. From a system modeling perspective, the use of data-driven models can be beneficial in cases where experimental data or high-fidelity data from more complex models are available, and a low computational cost with an acceptable accuracy is desired. The aim of this work is to investigate the capabilities of data-driven models based on two neural networks for modeling an open sorption reactor. The model takes as inputs the inlet temperature and sorbate concentration, and gives as outputs the reactor state of charge (SOC) and outlet temperature (TOUT). To account also for the thermal inertia of heat taking place in the reactor, both outputs are estimated with nonlinear autoregressive neural networks with exogenous inputs (NARXn), which account for the past n model outputs to determine the next output. Three neural network models are analyzed and several test cases are investigated to compare the performance of these neural network models with a high-fidelity CFD model. The results show that, for the SOC estimation, the NARX10 mean squared error (MSE) with respect to the high-fidelity CFD model was approximately two orders of magnitude smaller compared to the NARX1 MSE, resulting in a higher prediction accuracy. On the other hand, using the NARX10 architecture also for the TOUT estimation decreased the accuracy of TOUT estimations compared to a simpler FFNN neural network architecture considered in this work.