Improving accuracy of energy system models for an efficient energy transition: basis-oriented aggregation and machine learning

This work presents the results of applying a recently developed a-posteriori method called BasisOriented time series aggregation to real-world models. It extends it by removing the need for a dualsolution of the full PSOM and using machine learning techniques on previous runs to identifyapproximate partitions of the input space of a model with an unknown solution. The advantage ofthe Basis-Oriented approach is that it has been proven, mathematically, to be an exact aggregationof the full PSOM, measured as the difference between the aggregated and full models' objectivefunction values. By extending it through machine learning, we address the main limitation of its use:requiring a dual solution and relying on previous runs. This extension makes the Basis-Orientedapproach suitable to improve the computational tractability of operational models, which, despite notbeing as challenging as expansion ones (i.e., Mixed Integer Programs), still pose challenges tomodern computing platforms.