JuliaCon Local Paris 2025

Clustering is a popular tool for reducing the size of datasets before solving large-scale optimization problems—particularly in energy systems modeling. Common methods like k-means and k-medoids select points near the center of clusters, aiming to represent average behavior in the data. These centroids are then used as representative data points to build the reduced model. However, this approach often misses the points that matter most: the extremes that are more likely to be constraint-binding in the full optimization problem.

TulipaClustering.jl, takes a different approach: a (greedy) hull clustering algorithm tailored to reductions of optimization problems. Instead of maximizing intra-cluster similarity, we prioritize covering the convex hull of the input data with a small set of representative points. These representatives are more likely to approximate the binding regions of the original problem’s feasible space, leading to tighter approximations and better decisions downstream. Additionally, we use blended weights to approximate interior points as mixtures of the representatives. This allows the reduction to preserve key structural features of the original problem, while significantly reducing dimensionality.

In this talk, we describe the motivation, algorithm, and implementation behind TulipaClustering.jl, and show how it integrates into energy system models. We share practical results from energy-system optimization case studies, where greedy hull clustering outperforms conventional clustering methods in downstream optimization quality without sacrificing runtime. For energy modelers and optimization practitioners, TulipaClustering.jl bridges the gap between statistical model reduction and optimization-ready problem formulation.