Modeling the performance of a sorption thermal energy storage reactor using artificial neural networks

Sorption technology has the potential to provide high energy density thermal storage units with negligible losses. However, major experimental and computational advancements are necessary to unlock the full potential of such storage technology, and to efficiently model its performance at system scale. This work addresses for the first time, the development, use and capabilities of neural networks models to predict the performance of a sorption thermal energy storage system. This type of models has the potential to have a lower computational cost compared to traditional physics-based models and an easier integrability into broader energy system models. Two neural network architectures are proposed to predict dynamically the state of charge, outlet temperature and therefore thermal power output of a sorption storage reactor. Every neural network architecture has been investigated in 32 different configurations for the two operating modes (hydration and dehydration), and a systematic training procedure identified the best configuration for each architecture and each operating mode. A campaign of test cases was thoroughly investigated to assess the performance of the proposed neural network architectures. The results show that the proposed model is capable to accurately replicate and predict the dynamic behavior of the storage system, with mean squared error estimators below 2 · 10−3 and 50 °C2 for the state of charge and the outlet temperature outputs, respectively. Our findings, therefore, highlight the potential of an artificial neural networks based modelling technique for sorption heat storage, which is accurate, computationally efficient, and with the potential to be driven by real time data.