Skip to content

COSMO-REA6 Percentile Representative Year Methodology

1. Overview

For each of the 824 × 848 spatial cells, this algorithm identifies which calendar year from the COSMO-REA6 archive best represents the 10th-percentile (P10), median (P50), and 90th-percentile (P90) of the long-term solar radiation climate. The algorithm works over whatever years are actually present — the real production run used the full 1995–2019 archive (298 monthly files, not a clean 24-year boundary; DWD's real coverage stops partway through 2019).

The ranking metric is GHI (Global Horizontal Irradiance), the primary solar energy resource variable. Once the representative year is selected per cell per percentile, all variables from that year's file are carried into the output mosaic — not just GHI.

2. Why GHI as the Ranking Metric

GHI integrates both the direct-beam and diffuse solar components and is the single best indicator of PV/solar-thermal yield, building cooling load from solar gain, and daylight availability. This aligns with IEC 61724-1 and the ASHRAE TMY3 methodology, which ranks years by cumulative monthly global radiation.

3. Algorithm (per spatial cell)

3.1 Inputs

Input Shape Description
Monthly NetCDF files real production run: 298 COSMO_REA6_YYYY_MM_all_attrs.nc
Analysis period 1995–2019 ~24.8 years (real archive; see note above)
Spatial grid 824 × 848 COSMO-REA6 rotated-pole
Ranking metric daily GHI sum Summed per calendar day, per cell

Leap-year days (29 Feb) are removed before any calculation.

3.2 Steps

For each cell (i, j) and each month m:

1. Compute daily total GHI for every day in month m, for all N years.

2. Sum each year's daily totals into one cumulative monthly
   radiation figure:
       total[y,i,j] = sum of year y's daily GHI totals for month m

3. Take the target radiation level for each percentile ACROSS years:
       tgt_P10[i,j] = 10th percentile of total[:,i,j]
       tgt_P50[i,j] = 50th percentile (median)
       tgt_P90[i,j] = 90th percentile

4. Select the year sitting nearest each target level:
       best_P10[i,j] = year  with  min |total[y,i,j] - tgt_P10[i,j]|
       best_P50[i,j] = year  with  min |total[y,i,j] - tgt_P50[i,j]|
       best_P90[i,j] = year  with  min |total[y,i,j] - tgt_P90[i,j]|

This ranks candidate years by cumulative monthly global radiation, as in ASHRAE TMY3 (see section 2). Because total is a real-valued sum rather than a day count, the selection is effectively tie-free and the chosen year's brightness rank comes out at approximately the requested percentile.

3.2.1 Superseded selection rule (fixed 2026-08-19)

Until 2026-08-19 steps 2-4 instead pooled every year's daily values together, took the pooled P10/P50/P90 as thresholds, and picked the year minimising |fraction of that year's days below the threshold - q|.

That rule cannot produce the levels section 3.3 promises. A typical year has ~10% of its days below the pooled P10 — that is what the 10th percentile means — so min |cdf_P10 - 0.10| selected the typical year and actively rejected genuinely cloudy ones; min |cdf_P90 - 0.90| rejected sunny ones the same way. All three levels therefore chased the same target. Measured against the real 76-year ERA5-Land archive, the selected years' brightness ranks were 0.273 / 0.240 / 0.268 for P10 / P50 / P90 instead of 0.10 / 0.50 / 0.90 — P90 was returning years cloudier than the median.

The statistic was also a day count divided by max_days, so it could only take max_days + 1 distinct values. With 31-day months and 76 candidate years every cell had ties (median 12-20 years, mean ~42), and np.argmin awarded each tie to whichever year sorted first — handing the earliest year in the archive 52-59% of all cells at every level.

All three providers shared the defect; all three were fixed together, so any percentile output generated before 2026-08-19 should be regenerated.

3.3 Physical Interpretation

Output GHI level Interpretation
P10 10th percentile Extreme cloudy / low-solar year
P50 Median Typical Meteorological Year (TMY)
P90 90th percentile Extreme sunny / high-solar year

Adjacent cells can and do select different years — each cell optimises independently.

3.4 Mosaic Output

Because each cell independently selects its representative year, the output files are spatial mosaics:

P50 output (8760 h × 824 × 848):
  cell(0,0)     → all variables from year 2007
  cell(0,1)     → all variables from year 2003
  cell(823,847) → all variables from year 2011
  ...

The source_year(rlat, rlon) variable in each output file records the origin year for every cell.

4. Time Axis

All output files use a standard 8760-hour axis (365 days × 24 h). Leap-year files (8784 h) have their last 24 h (31 Dec hours 00–23) truncated to match.

5. Output Files

36 files total: 12 months × 3 percentile levels.

Pattern Percentile Content
cosmo_rea6_p10_MM_all_attrs.nc P10 Extreme low-GHI (cloudy) year
cosmo_rea6_p50_MM_all_attrs.nc P50 Median / typical year
cosmo_rea6_p90_MM_all_attrs.nc P90 Extreme high-GHI (sunny) year

Format: NetCDF-4 / HDF5, zlib compression level 1, float32.
Dimensions: time=8760, rlat=824, rlon=848.
Variables: T, GHI, DHI, WS_10M, PS, H_SNOW, SNOW_GSP, SNOW_CON (+ DNI if present), source_year.

6. Practical Notes

  • Re-runs: Existing valid output files are skipped automatically. Run with --clean to remove all output and force a full re-run.
  • Sample size: N = 24 years gives moderate percentile uncertainty, particularly at P10/P90, where the target level is interpolated from only two or three bracketing years. Interpret the tails with caution; ERA5-Land's 76-year archive is correspondingly tighter.
  • Time axis: monthly source files are not guaranteed to share one time axis — ERA5-Land's 1950-01 has no 00:00 stamp and so carries 743 hours starting at 01:00. The mosaic is sized from the longest axis across the winning years and each year is written at its own hour offset, so a short year lands in the right slots instead of shifting an hour early.
  • GHI-only ranking: Cells with uniformly low GHI (heavily clouded) may show inconsistent temperature or wind rankings relative to the selected P-level. Multi-variable ranking is a planned extension.