A LSTM-based Monte Carlo quantile regression method for satellite temperature prediction and data uncertainty quantification

Satellite on-orbit temperature data analysis is of great importance for improving satellite reliability and executing in-time health management. Due to the time series property of the on-orbit temperature data, the Long Short Term Memory (LSTM) network is a major method to solve the future temperature prediction problem. However, the data uncertainty caused by temperature noise is not considered in the traditional LSTM. To address this problem and improve prediction accuracy, a LSTM-based Monte Carlo Quantile Regression method (LSTM-MCQR) is proposed in this paper. On one hand, quantile levels sampled randomly from the uniform distribution are integrated with the time serial data as the input of the LSTM network in each training epoch. On the other hand, the data uncertainty is quantified through the trained model by sampling quantile levels again during the prediction process. The effectiveness of the proposed method is validated by an actual engineering case. Results demonstrate that a more accurate predicted temperature with the quantified uncertainty can be acquired, which is conducive to further reliability analysis.