Machine learning for predicting thermal power consumption of the Mars Express Spacecraft

Matej Petkovic, Redouane Boumghar, Martin Breskvar, Saso Dzeroski, Dragi Kocev, Jurica Levatic, Luke Lucas, Aljaz Osojnik, Bernard Zenko, Nikola SimidjievskiarXiv 2018 · DOI · OpenReview

TL;DR

This paper presents an improved machine learning pipeline for predicting the thermal power consumption of the Mars Express spacecraft, achieving a better balance between predictive accuracy and computational efficiency. By exploring different data granularities and advanced ensemble learning methods, the authors significantly reduce training time while maintaining high performance, offering valuable insights for spacecraft operations.

What Problem Does This Paper Address?

Spacecraft like Mars Express (MEX) have complex thermal management systems that consume a significant portion of their power. Accurately predicting this thermal power consumption (TPC) is crucial for optimizing scientific operations, especially as spacecraft age and their components degrade. Current methods rely on simplified models and expert knowledge, which can be insufficient for the dynamic and complex nature of TPC. This paper aims to develop more accurate and efficient predictive models using machine learning, building upon previous work that won a data mining competition but was computationally intensive.

Methods

The core of this work is an enhanced machine learning pipeline that addresses two key areas:

  • Data Granularity: The raw telemetry data was transformed into datasets with varying time resolutions (from 1 to 60 minutes). This coarser granularity significantly reduces the dataset size, leading to faster training times.
  • Machine Learning Algorithms: The authors explored three state-of-the-art ensemble learning methods for multi-target regression:
  • Local Random Forest of Predictive Clustering Trees (L-RF): Predicts each of the 33 thermal consumers individually.
  • Global Random Forest of Predictive Clustering Trees (G-RF): Predicts all 33 thermal consumers simultaneously with a single model.
  • XGBoost (XGB): A highly efficient implementation of Stochastic Gradient Boosting, also used for multi-target regression.

Additionally, the paper details sophisticated feature engineering techniques to extract informative features from raw telemetry data, including energy influx, command logs (DMOP), and flight dynamics events (FTL).

Datasets

The study utilizes data provided by the European Space Agency (ESA) for the Mars Express spacecraft. This dataset consists of:

  • Training Data (PTRAIN): Raw telemetry data from three Martian years (August 22, 2008 to April 14, 2014).
  • Test Data (DTEST): Raw telemetry data from the subsequent Martian year (April 14, 2014 to March 1, 2016).

The data includes measurements of electric current for 33 thermal power consumers and various context data such as solar aspect angles, mission operation plans, flight dynamics events, and long-term data.

Key Insights and Ideas

  • Coarser data granularity is a powerful lever for efficiency: Reducing data resolution from 1 minute to 15 or 30 minutes significantly cuts down training time with minimal loss in predictive performance.
  • XGBoost offers a compelling balance: While L-RF and G-RF are also effective, XGBoost consistently provides a strong trade-off between predictive accuracy and computational efficiency, especially with moderate data granularity.
  • Global vs. Local Models: Global models (G-RF) are faster but generally less accurate than local models (L-RF, XGB) which model each target separately. However, the difference in performance between local methods and G-RF is not always substantial.
  • Ensemble of Ensembles (EoE) for peak performance: Combining predictions from different well-performing ensembles (e.g., XGB and L-RF) can yield the highest accuracy, though at an increased computational cost.
  • DMOP features are critical: The Detailed Mission Operations Plans (DMOP) data, which logs spacecraft commands, plays a significant role in predicting thermal power consumption, particularly for XGB models.

Key Results

  • Time Efficiency: Learning times decrease substantially with increasing data granularity. For instance, using a 60-minute granularity dataset, L-RF, XGB, and G-RF take approximately 330, 4.1, and 2.3 hours, respectively, a significant improvement over the ~22,000 hours for L-RF with 1-minute granularity.
  • Predictive Performance:
    • XGB generally performs best or comparably to L-RF across various granularities, especially with moderate granularity (Δt ∈ {15, 30}).
  • G-RF is the most efficient but typically shows lower predictive performance.
  • An Ensemble of Ensembles (EoE) combining XGB and L-RF achieved the best overall performance with RMSE = 0.07779 on the Δt=15 dataset.
  • Feature Importance: DMOP features are found to be the most influential for XGB models, while energy influx features also play a significant role for RF ensembles.

Limitations and Future Work

  • DMOP data limitations: DMOP information is only available after a command is executed and its effect measured, limiting its use for predicting very long time horizons. Omitting DMOP data can severely impact model performance.
  • Future Work:
    • Further optimization of constructed features and exploration of new feature engineering approaches.
  • Applying the proposed methodology to other spacecraft subsystems or different spacecraft (e.g., XMM Newton, Integral, ExoMars) and rovers.
  • Investigating different approaches for engineering new informative features.

Summary updated

Abstract

The thermal subsystem of the Mars Express (MEX) spacecraft keeps the on-board equipment within its pre-defined operating temperatures range. To plan and optimize the scientific operations of MEX, its operators need to estimate in advance, as accurately as possible, the power consumption of the thermal subsystem. The remaining power can then be allocated for scientific purposes. We present a machine learning pipeline for efficiently constructing accurate predictive models for predicting the power of the thermal subsystem on board MEX. In particular, we employ state-of-the-art feature engineering approaches for transforming raw telemetry data, in turn used for constructing accurate models with different state-of-the-art machine learning methods. We show that the proposed pipeline considerably improve our previous (competition-winning) work in terms of time efficiency and predictive performance. Moreover, while achieving superior predictive performance, the constructed models also provide important insight into the spacecraft's behavior, allowing for further analyses and optimal planning of MEX's operation.
BibTeX
@misc{petkovic2018machine,
  author = {Matej Petkovic and Redouane Boumghar and Martin Breskvar and Saso Dzeroski and Dragi Kocev and Jurica Levatic and Luke Lucas and Aljaz Osojnik and Bernard Zenko and Nikola Simidjievski},
  title = {{Machine learning for predicting thermal power consumption of the Mars Express Spacecraft}},
  year = {2018},
  eprint = {1809.00542},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1809.00542},
}
Figure