Machine Learning for Spacecraft Operations Support - The Mars Express Power Challenge

Mars Express (MEX) has been orbiting Mars, generating great science for over 13 years. The aging spacecraft faces challenges; eclipse seasons are getting longer, eclipse durations are increasing and battery degradation is worsening. Accurate power modelling, respecting the power budget, becomes more vital as this ensures MEX health and maximum science. Empirical thermal power models provide reliable long term predictions but lack sensitivity. Telemetry data accumulated during the mission is a rich information source from which to derive a new model. This paper shows how the MEX Flight Control Team released 3 Martian years of data and reached out to Machine Learning (ML) enthusiasts asking them to predict a fourth year of spacecraft telemetry. In return, the participants were invited to an open data day to meet the MEX operators, tour ESA and find out about many exciting missions, and present and share their solutions with other candidates and ESA staff members. But the reward of solving a complex challenge, in a novel environment for space application, lead an incredible response from the worldwide ML community. Using open source solutions, candidates built data-driven models which have been able to predict the power consumption of 33 thermal lines, every hour, over a full Martian year (687 Earth days). These models improve sensitivity and accuracy by an order of magnitude. The number of scientific observations may increase even during power constrained periods when incorporating such models into MEX mission planning. By releasing spacecraft data and engaging with ML communities, ESA has gained a novel means to better exploit spacecraftresources, increase scientific return, and so, prolong mission life.