Understanding the performance impact of a massively parallel solver for energy system optimization models - a computational experiment using the PIPS-IPM++ solver for REMix instances

Kurzfassung

The complexity in the design of future integrated energy systems is reflected in the modeling tools used to analyze the interactions and synergies between energy technologies. As real-world systems become more interconnected, the complexity of model representations increases, resulting in a greater computational burden to obtain optimal solutions. Several approaches can address this challenge, including making trade-offs between model dimensions, using methods to reduce complexity, performing mathematical decomposition, and applying massively parallel solvers. While most of these approaches have been extensively studied, the application of massively parallel solvers has barely been explored due to their novelty. The main advantage of this approach is that it is well-suited to take advantage of modern high-performance computing infrastructure. Therefore, a more systematic evaluation of the types of model instances and decomposition strategies that can benefit from massively parallel solvers remains necessary. In this study, we identify capacity expansion decisions as the primary driver of computational complexity, particularly within the REMix framework, and demonstrate how to effectively leverage the underlying problem structure of these models. In a computational experiment we evaluate the performance of PIPS-IPM++ against a state-of-the-art interior-point solver. The results show a significant reduction in the total required wallclock time of about one order of magnitude, as well as a reduction in required computational resources. Our findings provide modelers with the necessary methods and capabilities to solve previously intractable, large-scale problems, thereby increasing the level of detail and explanatory power of energy system optimization models.