Advanced Feature Engineering and Prediction Analysis for Open Data of Mars Express Challenge

Gagan Manjunatha Gowda

Mastersthesis, Technische Universität Kaiserslautern, 6/2017.


Predictive analytics has attracted a lot of interest for a few years now and is particularly in the spotlight at the moment. Increasingly, organizations are turning to predictive analytics to up their value proposition and contribute towards scientific advancements.Crowdsourcing approaches like Kaggle and Kelvin is based on the fact that multiple approaches are utilized for a predictive modelling task. The thesis addresses the importance of feature engineering and machine learning based regression techniques in the context of Mars Express Power Challenge. The Mars Express Power Challenge, is a prediction analysis competition hosted by ESA. The task is to analyze Mars Express data collected over a period of three Martian years, which includes context data (influencing variables) and measurements of electric currents (target variables) with the goal of predicting average electric current of 33 thermal power lines or the fourth Martian year. This work successfully proposes one of the best prediction models for this task, also proposing a framework to approach prediction problems and selecting the best features influencing the target variables. An extensive range of models were considered, resulting in an ensemble model with KerasRegressor and Random Forest being the best. This model was ranked on the 8th spot in the competition. The final model is built with fine tuning of the above model with better feature engineering process and model design with XGBoost which results in the best prediction model. The influencing features of the thermal power consumption are then used by mission operators to generate better plans for power consumption, hence achieving safer flights and optimized scientific operations.


Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence