The Added Value of Advanced Feature Engineering and Selection for Machine Learning Models in Spacecraft Behavior Prediction

Abstract

This paper describes the approach of one of the top ranked prediction models at the Mars Express Power Challenge. Advanced feature engineering methods and information mining from the Mars Express Orbiter open data constitute an important step during which domain knowledge is incorporated. The available data describes the thermal subsystem power consumption and the operational context of the Mars Express Orbiter. The power produced by the solar panels and the one consumed by the orbiter’s platform are well known by operators, as opposed to the power consumption of the thermal subsystem which reacts to keep subsystems at a given range of working temperatures. The residual power is then used for scientific observation. This paper presents an iterative and interactive pipeline framework which uses machine learning to predict, with more accuracy, the thermal power consumption. The prediction model, along with the estimation of the thermal power consumption, also provides insight into the effect of the context which could help operators to exploit spacecraft resources, thereby prolonged mission life.

Appendix

Table 4 shows the list of important features extracted from FEAT model which also signifies the correctness of the predicted result.

Table 4 Influencing features for thermal power consumption of Mars Express Spacecraft