Contribution type: lightning talk and poster
Data-driven sector reduction using differentiable optimization layers
Presenter: Tim Graulich
Description:
Energy system models are becoming increasingly complex as they aim to represent additional system dynamics at higher spatial and temporal resolutions. A key driver of this trend is the growing importance of sector coupling, which requires the simultaneous modeling of multiple sectors. This substantially increases model size and solution times, creating significant barriers to adopt multi-sectoral studies. In particular, when stakeholders require high resolution and uncertainty quantification in a specific sector, extending the same level of detail to additional sectors often becomes computationally infeasible. Thus, a simpler and computationally lighter way is required to include additional sectors in energy system models. One common strategy is spatially aggregate nodes in the additional sectors. However, aggregating models generally creates errors between the full and aggregated model, thus affecting the model outcomes. Although previous work has investigated how to reduce these errors, such as through additional constraints or iterative resolution refinement, no existing methods explicitly account for predefined accuracy measurements while optimizing the aggregation itself to minimize induced errors. In this study we propose a data-driven workflow that utilizes parametrized aggregation models and differentiable optimization layers, to minimize case-oriented metrics between the full and aggregated model. In a case study, we develop a reduced spatial representation of the European H2 sector, while trying to keep electrolyzer operation as close as possible to the original system. Using congestion aware clustering and learned bounds on the aggregated system, we minimize the error and compare it to naive clustering approaches.
Background:
This PhD project is hosted at DTU and is part of the MuESSLi project within Cresym; involving 5 PhDs from DTU, TU Delft and UP Comillas. It is funded by RTE, Total Energies and NaTran.
Optional links:
- MuESSLi project: https://cresym.eu/muessli/
- Repo will be added later