Using physics-based regularization in Artificial Neural Networks to predict thermochemical energy storage systems

The shift towards renewable energy poses urgency for the development of energy storage systems. One promising technology is a thermochemical energy storage device using a reversible endothermic/exothermic reaction of CaO/Ca(OH)₂. The system can be modelled as a non-isothermal, single-phase gas, multi-component porous medium flow problem. Due to the complex non-linear nature of the problem, numerical simulation is expensive, and therefore, surrogate models are required for accurate and real-time-capable prediction of the system. One solution is to use Artificial Neural Networks (ANN). However, a fully black-box model often produces physically unrealistic predictions. Our work addresses this issue by introducing a regularization term in the loss function based on mole and energy balance to train the ANN to realize extensive physical relationships between the whole set of inputs and outputs. This term penalizes the network if the mole and/or energy balance is not fulfilled. We use 100 training datasets obtained from simulation using a numerical model with additional white noise (as a replacement of observation data). Both the conventional and the physics-regularized ANN are then trained using these datasets, using the Levenberg-Marquardt algorithm combined with Bayesian Regularization. The results show that additional physics-based regularization does not improve the ANN test performance based on Mean Squared Error criterion, but more importantly it removes the physically unrealistic results and improves the stability of the predictions. Therefore, this work serves as a promising first step towards an improved ANN for predicting the thermochemical energy storage system.