Bridging the climate to energy data gap: simulated annealing for representative climate year selection

TL;DR

This paper introduces a simulated annealing approach to select representative climate years for energy system modeling, significantly improving the representativeness over current practices.

Contribution

It proposes a novel optimization method using simulated annealing with a Wasserstein distance metric to improve climate year selection for energy models.

Findings

Simulated annealing outperforms other methods in representativeness.

Selected climate subsets are 2.5-3.5 times more representative than current practices.

The method achieves an effective sample size four to five times the subset size.

Abstract

Energy system models are increasingly dependent on representative climate input. Yet, a fundamental mismatch persists between the hundreds of simulated years often used in climate science and the handful of years that computationally demanding power system models can process. Current practice, including ENTSO-E's European Resource Adequacy Assessment, relies on climate year selections that have not been validated against explicit representativeness criteria. This risks biased investment decisions and blind spots for plausible weather conditions. This study proposes simulated annealing as an optimisation method for selecting representative subsets of complete climate years from large climate ensembles. Representativeness is quantified using the seasonal sliced Wasserstein distance, a metric from optimal transport theory that captures representativeness on marginal distributions, inter-variable correlations, and seasonal structure simultaneously. We evaluate simulated annealing against the alternative methods random search, filtered random search, and K-Medoids clustering across three test cases spanning the Netherlands and Europe, using 180 climate years from the Pan-European Climate Database as a reference. Simulated annealing consistently produces the most representative subsets and outperforms all compared methods. Simulated annealing achieves an effective sample size four to five times the actual subset size. The resulting subsets are roughly 2.5--3.5 times more representative than current ENTSO-E practice. The method is application-agnostic and its output can serve as a validated climate data input to any subsequent (energy) impact study.