Machine Learning for Time Series: Space Applications

In this section, machine learning techniques discussed in Machine Learning for Time Series are applied in the context of Space Traffic Management.

Applications

ESA Solar Activity

Investigation into predictive power of machine learning on solar activity, geomagnetic activity, and consequently atmospheric density. The variables to learn are:

• F10: Solar radio flux (top figure)
• Ap: Average level of geomagnetic activity (bottom figure)

Evolution of daily F10 [solar flux units].

Evolution of daily Ap.

Estimation of Collision Risk

Use time series machine learning to learn how to estimate the risk of collisions between two objects in space (the chaser and the target), based on the information contained in successive Conjuction Data Messages (CDMs).

Evolution of collision risk between two objects with time to their closest approach.

Each CDM contains multiple variables, such as:

• The position and covariance matrices of the two objects
• The miss distance

Therefore, the task must be tackled with multivariate time series techniques.

Evolution of covariance measures with time to closest approach.

The estimation of risk can be approached from different perspectives:

• As a time series forecasting task, where the future values of the collision are predicted based on the past ones.
• As a time series classification task, where the situation at Time of Closest Approach (TCA) between two objects is classified either as high-risk or low-risk, based on the evolution of the two objects days before TCA.
• As a time series regression task, where the exact risk at one specific CDM (e.g. the last one) must be predicted based on values from previous CDMs.

This task forms part of the ESA collision avoidance challenge.