Resource Forecasting for Satellite Operations using Multivariate Time Series Data
Most of the resources of a satellite, once launched into space, are very limited and have to be shared with all on-board consumers. For example, the current power is mostly produced by small solar panels. The power output of these panels restrict the on-line time of power consuming equipment like cameras or radio devices. Therefore, one main job of satellites operators is to plan the tasks of a satellite in respect to its available resources. As real-time data is only available for a few minutes every couple of hours for most LEO (Low Earth Orbit) satellites, a dynamic and generic means of prediction of resources could greatly aid operations.
The German Space Operations Center (GSOC) at the German Aerospace Center (DLR) commands a fleet of modern satellites, each with thousands of on-board sensors to keep track of satellites system status. Many of the satellites are in routine-phase, allowing us to accumulate a sufficiently large training set for machine learning purposes.
The goal of this project is to forecast a single parameter, e.g., a battery voltage, using all related parameters and planned commands which are to be executed on the satellite as input. The forecast horizon should at least be a couple of hours. This means, to investigate and compare different methods for multivariate time series forecasting, i.e., supervised/semi-supervised ML techniques such as regression, neural networks, random forests, etc. The result is an improved resource planning with greater efficiency and a possible enhancement of the current mission planning algorithms. Launch your mission with us!