An Improved Uncertainty Calibration Approach for Reliable Satellite Temperature Interval Prediction

Precise forecasting of satellite temperature interval prediction is essential for assessing the reliability and health of satellites. Yet, traditional methods for quantifying uncertainty often fall short when it comes to the intricate challenge of satellite temperature prediction, leading to less accurate estimates. Consequently, there is a pressing need to refine the accuracy of uncertainty quantification by readjusting the uncertainty outcomes. This paper presents a sophisticated approach to calibrating model uncertainty for satellite temperature forecasting, utilizing Monte Carlo dropout (MCD) and quantile calibration techniques. The MCD technique is used to estimate model uncertainty by generating samples of the output distribution. Simultaneously, the quality of uncertainty estimation is improved by fine-tuning the quantile levels through quantile calibration, leading to a more precisely calibrated forecast interval. The effectiveness of this proposed methodology is confirmed through two simulation scenarios and a practical engineering application, showing that the calibrated forecast interval's coverage probability is more closely aligned with the desired confidence levels, thus enhancing the credibility of the projected uncertainty range.