Transient system-wide surrogate model for spacecraft thermal analysis using recurrent neural networks with POD data reduction

Abstract

Spacecraft operations are highly uncertain and exposed to extremely harsh thermal environments. Therefore, it is essential to design a spacecraft for contingencies using comprehensive thermal analysis. However, the conventional thermal mathematical model (TMM) is difficult to implement in many cases because of its high computational costs. As such, various surrogate models have been proposed to reduce computational costs. However, none of these models can predict the spatiotemporal temperature distribution of an entire spacecraft for various analysis cases, resulting in difficulties in completely replacing the TMM-based thermal analysis with low computational costs. This study proposes a surrogate model that can quickly estimate the transient temperature distribution of an entire system for various analysis cases. The proposed surrogate model, POD-RNN, can efficiently learn spatiotemporal features using a combination of proper orthogonal decomposition (POD) and recurrent neural network (RNN). Numerical experiments to evaluate the performance of the proposed method show that it achieves a high prediction accuracy with an average error of 1.53 K. The proposed method also successfully predicted the temperature distribution with a very low computational cost of 0.23% compared with the TMM thermal analysis. The proposed POD-RNN can quickly estimate the same amount of information as TMM-based thermal analysis, providing a more flexible application for thermal analysis than conventional surrogate models.