Welcome to the Project UNSEEN¶
Project goals and research questions¶
the development of robust ESM scenarios under uncertainty becomes possible, as robustness can be estimated by the frequency of certain scenario “families”
complex questions can be answered (using indicators from multiple models)
the high resolution with respect to many parameters allows very sophisticated analyses and the restrictions of previous model runs are overcome
what must be the nature of a suitable model in terms of accuracy and size with MIP formulation?
what neural network topology is most appropriate for this type of problem?
what is the design and implementation of the feedback system described above in an HPC environment?
What is UNSEEN¶
In the predecessor project BEAM-ME, considerable successes have already been achieved in reducing the computing time of optimizing energy system models, both through more efficient algorithms and the upgrading of previous program code for the use of high-speed computing clusters or High Performance Computing (HPC). In the UNSEEN project, this work is now to be consistently continued in order, on the one hand, to calculate millions of scenarios in relatively short periods of time and, on the other hand, to consistently further develop the solution algorithms. From a scientific point of view, the focus will be on a better validation of the results by parameter variations of the models and an increase of the resolution level of the models. In detail, very large parameter space scans create a data basis that does not yet exist in this form. Millions of model runs on supercomputers ensure very good coverage of almost all combinations of model input parameters that are subject to high uncertainties. These data are independent of further methodological development, i.e. they can be statistically analyzed for policy-relevant areas. The central approach to be followed in this research project for solving a large number of high-resolution instances of energy system models is the use of neural networks. The underlying idea in using a neural network, or better machine learning methods, is to quickly “predict” the outcome of an optimization problem. Such a prediction can be used in two ways. On the one hand, a solution determined in this way can be introduced into the optimization procedure as a starting solution, thus potentially significantly reducing the time required to find a solution. In the end, it does not matter whether the solution was admissible or inadmissible, i.e., a correct solution or one that violates the constraints of the model, since the optimization procedure uses only those parts of the predicted solution that are conducive to the solution. This method development should be well transferable to other optimization projects, since it is basically independent of the considered optimization problem. Moreover, by coupling the methods with security of supply and economic analyses as well as consistent ESM scenario requirements, it becomes possible to identify robust solutions that are near-optimal with respect to several objectives.
Acknowledgements¶
The UNSEEN project was funded by the German Federal Ministry for Economic Affairs and Energy under grant number FKZ 03EI1004A. The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V. (www.gauss-centre.eu) for funding this project by providing computing time through the John von Neumann Institute for Computing (NIC) on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre (JSC).