Decision-Based vs. Distribution-Driven Clustering for Stochastic Energy System Design Optimization

TL;DR

This paper compares decision-based and distribution-driven clustering methods for stochastic energy system design optimization, demonstrating their similar cost outcomes but differing computational efficiencies in a real-world case study.

Contribution

It introduces the first application of decision-based clustering to energy system design optimization and evaluates its effectiveness compared to traditional distribution-driven methods.

Findings

Both clustering methods produce similar cost-efficient designs.

Decision-based clustering requires more computational resources.

The study is based on a real-world industrial energy system case.

Abstract

Stochastic programming is widely used for energy system design optimization under uncertainty but can exponentially increase the computational complexity with the number of scenarios. Common scenario reduction techniques, like moments-matching or distribution-driven clustering, pre-select representative scenarios based on input parameters. In contrast, decision-based clustering groups scenarios by similarity in resulting model decisions. Decision-based clustering has shown potential in network design and fleet planning. However, its potential in energy system design remains unexplored. In our work, we examine the effectiveness of decision-based clustering in energy system design using a four-step method: 1) Determine the optimal design for each scenario; 2) Aggregate and normalize installed capacities as features reflecting optimal decisions; 3) Use these features for k-medoids clustering to identify representative scenarios; 4) Utilize these scenarios to optimize cost in stochastic programming. We apply our method to a real-world industrial energy system modeled as a mixed-integer linear program. We incorporate uncertainty by scaling time series with representative factors. We generate 500 single-year scenarios via Monte Carlo sampling, which we reduce using decision-based clustering. For benchmarking, we conduct distribution-driven k-medoids clustering based on the representative factors. In our case studies, both clustering methods yield designs with similar cost efficiency, although decision-based clustering requires substantially more computational resources. To our knowledge, this is the first application of decision-based clustering on energy system design optimization. Future research should investigate the conditions under which decision-based clustering yields more cost-efficient designs compared to distribution-driven clustering.