Physics-informed neural network for real-time thermal modeling of large-scale borehole thermal energy storage systems

To exploit the full potential of borehole thermal energy storage (BTES) systems, real-time predictive system performance modeling is required to enable heat extraction to match real-time heating demand. This study presents a data-driven modeling approach utilizing a physics-informed neural network (PINN), which can combine both the explanatory power of physical models and the expressive power of neural networks and compares it with a conventional neural network (NN). We utilized a dataset from a real-world BTES system in Chifeng, China, which included 11,947 h of continuous monitoring of fluid temperature, flow rate, and multiple soil temperature measurement points. After training, the PINN and NN achieved mean absolute errors in outlet temperature predictions of 0.3 °C and 0.6 °C, with R2 values of 0.996 and 0.984, respectively. The PINN demonstrated superior predictive accuracy compared with the conventional NN, and further experiments confirmed that the PINN exhibited robust training performance with less training data. We also assessed the impact of varying flow rates of the BTES heat transfer fluid on heat extraction, and the results highlighted the BTES system's ability to adapt to real-time changes in heating demand.