Selecting representative climate years for national to continental-scale energy system studies

Due to the ongoing energy transition to variable renewable energy sources, climate variability plays a central role in energy system studies. Climate science routinely addresses this variability by simulating large ensembles spanning hundreds to thousands of model years. However, energy system and power-grid models used for industrial applications are computationally intensive and typically cannot process more than a few years to a few decades of climate data. This mismatch necessitates the selection of a small but representative subset of climate years.

A common workaround is the use of composite or “typical” meteorological years constructed from individual months. While computationally efficient, such synthetic time series disrupt temporal coherence, and fail to capture memory effects that are critical for adequacy assessments, such as storage dynamics of hydropower. As a result, many energy system studies instead select a limited number of complete climate years, typically ranging from one to fifty. Selecting such subsets from large climate simulations constitutes a combinatorial optimisation problem: choosing X years from N>>X, for which brute-force optimisation is computationally infeasible due to ‘combinatorial explosion’.

Current practices rely heavily on (pseudo) random sampling or heuristic selection methods, including clustering-based approaches such as k-medoids (or k-means). While useful, these methods provide no guarantee of near-optimal solutions and often struggle to balance representativeness across multiple, interacting climate variables relevant for energy systems.

In this study, we systematically review existing climate-year selection methodologies and introduce simulated annealing as a flexible and computationally efficient optimisation framework for selecting representative subsets of complete climate years. The method targets representativeness of the joint distribution of multiple energy generation and demand variables. We apply the approach to the Pan-European Climate Database, which comprises 85 years of simulations from six CMIP6 climate models under four SSP scenarios, together with associated energy demand and renewable generation time series. Two use cases are considered: the selection of a larger subset of 30 representative years for adequacy-type studies, and a smaller subset of 5 years for investment-type studies. Across both cases and for both national and contintental-scale applications, simulated annealing consistently outperforms existing methods, proving to be the most robust method for climate year selection in large-scale energy system modelling.