Annex III: Weather Year Selection Methodology //

The methodology is based on weather and energy variables from the PECDv4.2 dataset (Annex II), featuring the following models:

The methodology uses:

For each study horizon (2030, 2035, 2040, 2050), a window of 10 years per climate model is used (e. g., TY2030: 2025–2034).

Together with the three climate models (CMR5, ECE3, MEHR) , this yields 30 candidate climate‑year series per target year, from which three representative years are selected.

1// Computation of annual averages and cumulative anomalies

2// Calculation of overall statistics (mean, standard deviation) and deltas

3// Application of regional weighting

4// Normalisation of all parameters

5// Dimensionality reduction via Principal Component Analysis (PCA)

6// Clustering using k‑means methodology (k = 3)

The method provides three representative weather years per TYNDP target year, each reflecting a distinct climate pattern (see Table 33).

Although the CMR5 model appears over‑represented in final selections, this has no climatic significance and does not bias the methodology.

Energy system planning studies must account for weather-driven variability in both demand (temperature-dependent heating and cooling) and supply (wind, solar, and hydro generation). Simulating the 30 scenarios from the three climate models for each target year is computationally heavy and time consuming. The Weather Scenario Selection methodology reduces this set to k = 3 representative weather years per target year, each carrying a probability weight that reflects the frequency of its weather characteristics. This reduction aims make the computational problem tractable while keeping an appropriate level of quality on the final outcomes.

The selected scenarios and their weights allow downstream market modelling processes to approximate expected values over weather uncertainty. This reduces the simulation burden from 30 to 3 runs per target year.

The Weather Scenario Selection is performed early in the scenario building process. Its outputs are the three representative weather scenarios and their associated weights are used as inputs to:

Three categories of weather-dependent data are loaded for each of the 30 candidate weather scenarios and 53 bidding zones:

Heating and Cooling Degree Days (HDD / CDD)

Hourly temperature data from the three climate models is converted to daily averages. A threshold transformation is then applied using zone-specific heating and cooling thresholds derived from econometric calibration (per-country GDP, building thermal transmittance, and regression coefficients). Temperatures above the cooling threshold produce cooling degrees for the definition of CDD; temperatures below the heating threshold produce heating degrees for the definition of HDD; temperatures within the comfort band produce zero contribution to both HDD and CDD.

Renewable Generation (Solar PV, Wind Onshore, Wind Offshore)

PECD capacity factor profiles are multiplied by installed generation capacity (MW) from the TYNDP generation capacity dataset for the target year. Sub-types are aggregated to technology level (e. g., Solar PV utility tracking, non-tracking, industrial rooftop, and residential rooftop are summed into a single “Solar” category; fixed and floating offshore wind are summed into the “Offshore” category).

Hydro Inflows (Run-of-River, Reservoir)

Inflow time series are loaded directly from PECD files for each zone, validated against installed hydro capacity in the generation dataset.

After the previous Stage 2, each of the 30 weather scenarios is described by a 12-dimensional feature vector comprising six normalised mean-anomaly values and six normalised variability-anomaly values, one per technology.

These 12 dimensions are organised into two independent subspaces: (i) the mean-anomaly matrix A (30 scenarios × 6 technologies), whose columns are the z-scored mean deltas for each technology; and (ii) the variability-anomaly matrix S (30 scenarios × 6 technologies), whose columns are the z-scored variability deltas.

A Principal Component Analysis (PCA) is applied independently to each subspace, retaining only the first principal component (PC1). Each weather scenario is then represented as a point in a two-dimensional space: the first axis (PC1 of the mean-anomaly matrix) captures the dominant mode of co-variation in mean generation / demand levels across technologies; the second axis (PC1 of the variability-anomaly matrix) captures the dominant mode of co-variation in intra-annual variability.

The 30 points in the reduced 2D space are partitioned into k = 3 clusters by k-means clustering, minimising the within-cluster sum of squares. The algorithm is run with 10 random initialisations to mitigate sensitivity to initial centroid placement, and the partition with the lowest total within-cluster variance is retained. A fixed random seed ensures full reproducibility.

The representative year for each cluster is the scenario closest to the cluster mean in the two-dimensional PCA space. By selecting an actual weather scenario rather than the centroid itself (which is a synthetic point with no corresponding time series), the representative weather scenario chosen always corresponds to an existing model year with a complete, physically consistent set of profiles across all technologies and zones.

The probability weight of each representative scenario equals the proportion of historical weather years assigned to its cluster: w_k = \lvert C_k \rvert / N, where \lvert C_k \rvert is the number of years in cluster k and N = 30 is the total number of candidate scenarios.

This is a non-parametric density estimator under the ergodicity assumption: the temporal frequency of weather patterns in a sufficiently long temporal record approximates their probability of occurrence. The weights are non-negative, sum to unity by construction, and represent the empirical likelihood of each weather regime — not a measure of weather severity or extremity.

The following Table 36 summarises the selected representative weather years and their probability weights for each target year. For each target year, the k-means algorithm identifies three distinct weather regimes.

One cluster typically contains roughly half of the 30 candidates, representing the most common weather pattern, while the other two clusters capture less frequent but distinct regimes.