Digital Twin of Power-to-Heat Latent Heat Thermal Energy Storage

Abstract

Latent heat storage thermal energy storage (LHTES) uses phase change materials (PCMs) to exploit the advantage of high energy density and the possibility of heat supply at relatively constant temperatures. Despite the advantages, the implementation of high-temperature PCM systems, particularly those exceeding 150°C, remains limited in industrial and commercial sectors. Previous studies at DLR demonstrated a high-temperature power-to-heat storage system at lab scale using a KNO₃-NaNO₃ eutectic mixture (melting point ~222°C) as the phase change material. The system achieved a storage capacity of 4.87 kWh, an average power output of 100 kW, and charging efficiencies between 65% and 90%, highlighting the need for improved monitoring and control strategies. The aim of this study is to develop digital twin models that can monitor the thermophysical state of the LHTES in high resolution, based on limited number of sensor data. Conventional modelling strategies are posing high computational cost due to multi-physical and transient nature of the phase change process. On the other hand, digital twins require real time modelling and monitoring of the physics. To achieve such a model, an approach that is combining the data-driven and physics-based modelling is combined. Previous work at DLR provides simulation (finite element method-based) and experimental data on wide range of operating conditions of the storage. These studies provide the training data for neural network which built the core of the digital twin model. Because of its availability and high resolution, synthetic data is prioritized on this study. Stored thermal energy can be calculated by two different references. First is the virtual sensors of temperature and the phase state of the PCM. The second is the total energy balance at the LHTES boundary, where the input and output heat data are available. An additional loss function that incorporates the difference between these two references should increase the accuracy of the model by accounting for the conservation of energy. The results of the study focus on the model with a 2-hidden-layer neural network, trained for 100 epochs with a batch size of 32 . It is shown that dynamics of the charging and discharging of the thermal system can be captured with an accuracy of 84% for temperature sensors and 91% for the phase state. Expanding this study with experimental data and modeling of heat transfer mechanisms will enhance the reliability and applicability of the digital twin model for LHTES systems. Future work of this study will focus on integrating real sensor data so that the model can be validated under practical operating conditions, enabling improved monitoring, control, and optimization of high-temperature thermal energy storage systems.