Open Access

Local and regional energy systems are becoming increasingly entangled. Therefore, models for optimizing these energy systems are becoming more and more complex and the required computing resources (run-time and random access memory usage) are increasing rapidly. The computational requirements can basically be reduced solver-based (mathematical optimization of the solving process) or model-based (simplification of the real-world problem in the model). This paper deals with identifying how the required computational requirements for solving optimization models of multi-energy systems with high spatial resolution change with increasing model complexity and which model-based approaches enable to reduce the requirements with the lowest possible model deviations. A total of 12 temporal model reductions (reduction of the number of modeled time steps), nine techno-spatial model reductions (reduction of possible solutions), and five combined reduction schemes were theoretically analyzed and practically applied to a test case. The improvement in reducing the usage of computational resources and the impact on the quality of the results were quantified by comparing the results with a non-simplified reference case. The results show, that the run-time to solve a model increases quadratically and memory usage increases linearly with increasing model complexity. The application of various model adaption methods have enabled a reduction of the run-time by over 99% and the memory usage by up to 88%. At the same time, however, some of the methods led to significant deviations of the model results. Other methods require a profound prior knowledge and understanding of the investigated energy systems to be applied. In order to reduce the run-time and memory requirements for investment optimization, while maintaining good quality results, we recommend the application of (1) a pre-model that is used to (1a) perform technological pre-selection and (1b) define reasonable technological boundaries, (2) spatial sub-modeling along network nodes, and 3) temporal simplification by only modeling every nth day (temporal slicing), where at least 20% of the original time steps are modeled. Further simplifications such as spatial clustering or larger temporal simplification can further reduce the computational effort, but also result in significant model deviations.