A data-driven reduced-order model based on TCN-GRU and POD for temperature field rapid prediction of space nuclear reactor

Reduced-order model (ROM) offers a powerful and efficient tool for design and optimization of complex systems by simplifying high-dimensional systems, which significantly reduces computational complexity while retaining critical physical features. The traditional ROM cannot accurately capture the detailed features of transient physical fields. In recent years, the booming development of temporal neural networks would be useful to overcome this shortcoming. In this work, a data-driven ROM was created based on the Temporal Convolutional Network (TCN) and Gated Recurrent Unit (GRU) with proper orthogonal decomposition (POD) approach. Then the established data-driven ROM was applied to the rapid prediction of three-dimensional transient temperature field in space lithium-cooled fast reactor core. With the POD approach, main features of the transient temperature field are preserved using only 2 or 3 modes. Then the hidden time-sequence variations of the feature coefficients were captured and trained using TCN-GRU model to realize the prediction of the feature coefficients at different moments. The three-dimensional temperature field of space lithium-cooled core can be quickly reconstructed by a linear combination of reduced-order basis and feature coefficients. By comparing the predicted results with Computational Fluid Dynamics (CFD) simulation results, the data-driven ROM was validated. The maximum absolute errors of temperature fields for the two transient cases are 4.11 K and 1.14 K, respectively, demonstrating that the developed data-driven ROM can accurately and rapidly predict the temperature field of the space lithium-cooled core at different moments. This work contributes a new measure for on-orbit transient operation analysis and provides a valuable reference for the construction of digital twins for space nuclear power.