Ensemble learning, an alternative to a whole energy model ? Replacing a linear programming formulated energy system with supervised learning tools
Files
Van_Zeebroeck_38651700_Van_der_Burght_11211700_2022.pdf
Closed access - Adobe PDF
- 5.22 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Decision makers often resort to whole energy models to help them get insights on possible energy transition strategies. EnergyScope Typical Days (ESTD) is one of those models. It allows to determine the yearly optimal technology mix at a regional scale by accounting for an hourly resolution. However, because the chosen strategy has a long term impact and that the model parameters are uncertain, such as the availability of the different resources, a possible strategy could become obsolete really quickly. Robust solutions are therefore being studied by adding uncertainties directly in the formulation of the model. However, adding uncertainty ranges to the different parameters comes at an important computational expense. Therefore this thesis explores the possibility of replacing the Linear Programming (LP) framework of ESTD through tree based regressions. This would allow a quick evaluation of different situations therefore enabling further research on robust solutions. To answer such problem, we provide a dataset based on uncertainty quantification as well as the implementation of a Random Forest and eXtreme Gradient Boosting algorithm. The predictions analysis will highlight the strength and limitations of our implementation.