ERA5: The new champion of wind power modelling?

Reanalysis provides a consistent reprocessing of meteorological observations using an unchanging data assimilation system, producing variables such as wind speed, temperature, and pressure that are widely used for wind power modelling. Reanalyses are attractive due to their global and mostly free availability, and are used for applications like generating long wind power time series for grid studies and for long-term correction (LTC) of pre-construction wind measurements. Historically, higher temporal and spatial resolutions have been introduced with new reanalyses. MERRA/MERRA-2 (NASA) have been popular due to hourly resolution and wind speeds at 50 m, with prior studies finding good performance for country-level wind power modelling. ERA5, released starting in 2017, brings hourly resolution, freely available 100 m wind speeds (suited to modern turbines), and higher spatial resolution (~31 km) than ERA-Interim and MERRA-2. However, higher resolution does not always guarantee better correlations with measurements. The main objective here is to compare MERRA-2 and ERA5 for modelling wind power, both at country scale (Germany, Denmark, France, Sweden, and the BPA area in NW USA) and at individual turbine scale (1051 Swedish turbines), and to propose a new metric for system size and dispersion.

Data: ERA5 (initial release covering 2010–2016; plan to extend to 1950–present) and MERRA-2 (near real-time updates, 50 m winds) were used. ERA5 has ~31 km grid spacing and 100 m winds; MERRA-2 uses a ~0.5°×0.625° grid and 50 m winds. Country-wise generation measurements, farm metadata (coordinates, capacities, commissioning/decommissioning), and installed capacity time series were sourced from OPSD (Germany, Denmark, France), BPA, and national datasets. For Sweden, individual turbine generation time series were obtained (Cesar database) for 1051 turbines (≥1 MW, commissioned 2000+, 1–6 years of data). For Swedish turbines lacking coordinates, locations were estimated by selecting the reanalysis grid point yielding the highest correlation between modelled and measured generation; validation on a subset showed negligible bias in correlations and typical location errors of 36 km (MERRA-2) and 15 km (ERA5), consistent with grid resolution.

Country-wise modelling of hourly generation: For each wind farm/turbine, reanalysis winds were bilinearly interpolated in the horizontal (vertical handled via model levels: 50 m for MERRA-2, 100 m for ERA5). A generic turbine power curve with specific rating 360 W/m² was used for all units and smoothed (“multi-turbine” curve) with a Gaussian kernel (standard deviation 1.5 m/s) to represent farm aggregation. To emulate losses, two components were used: internal losses (e.g., wake, mild icing, blade degradation) modelled as a 10% reduction of available wind energy, and external losses/country-level smoothing (to prevent unrealistically frequent near-1 p.u. outputs due to under-dispersion of reanalysis winds) implemented as a multiplicative factor equal to 101% of the measured maximum country-wise output. Because reanalyses are not suitable for absolute mean wind speeds at specific sites and hub heights were unavailable, mean wind speeds per farm were adjusted via a capacity factor (CF) target: CF_i = CF_country + β (u_i − ū_country), with β = 1/40 and u_i, ū_country the reanalysis mean winds at model heights for the site and the capacity-weighted country mean, respectively. Wind speed time series were then linearly scaled to achieve the desired CFs (assuming constant shear). For Denmark and Germany, onshore and offshore fleets were treated separately. Model outputs and measurements were normalised by their respective installed capacity time series to p.u. values.

Individual turbine modelling (Sweden): Internal losses were set to 15% and external losses to 0% (as single turbines can reach 1 p.u.), and the smoothing parameter was reduced to 0.5 m/s to better represent individual turbine variability. The modelling ensures correct average CF and maxima by construction via the use of measured CF and capacity normalisation.

Downtime detection and filtering: To avoid penalising reanalyses for periods of turbine downtime (maintenance/faults), an automated detector flagged periods with zero measured output lasting at least 5 hours. A zero-output period was labelled as downtime if any of: (1) duration > 1 week; (2) duration ≥ 12 hours and modelled mean generation ≥ 0.05 p.u.; (3) modelled mean generation ≥ 0.15 p.u. Validation indicated ~9% false negatives and ~8% false positives relative to true downtime hours. After removing downtime samples, reanalysis-based generation was recomputed based on the updated (higher) measured CF.

Long-term correction (LTC): For each of 893 turbines (having ≥2 years of data), one randomly chosen year served as the baseline. ERA5- and MERRA-2-based generation were modelled for the full measurement span, with mean wind speeds adjusted so that model CF during the baseline period matched the measured CF. Predicted CF for the remaining 1–5 years (evaluation period) was the average modelled generation over that period. Performance was compared to a naïve estimator assuming the evaluation CF equals the baseline CF.

Equivalent system radius (req): A new metric quantifies the effective size/dispersion of a wind power system as the radius of a disk of uniformly distributed capacity that yields the same aggregate variance as the actual fleet, assuming inter-site correlation decays exponentially with separation distance (ρ = e^(−d/D)). Starting from Var(P_tot) = σ² Σ_i Σ_j P_i P_j ρ_ij for the real fleet (P_i capacity share; σ standard deviation per-unit at sites), and equating this to the continuous double integral over a uniform disk of radius req, the metric is solved numerically using D = 500 km, a commonly reported correlation length scale. For practical use, a polynomial approximation relates req to z = Σ_i Σ_j P_i P_j e^(−d_ij/D) over the range 0.22 ≤ z ≤ 1 with small fitting error (~0.5 km).

Country-wise modelling (hourly): Across five systems (Germany, Denmark, France, Sweden, BPA), ERA5 outperformed MERRA-2 in all assessed aspects, with average RMSE reduced by ~22% and MAE by ~24%. Highlights from quantitative results (measured vs modelled p.u. time series):

• Germany: Correlation increased 0.982→0.987; RMSE 2.82%→2.35%; MAE 2.06%→1.66%; step-change σ error +12.8%→+7.5%.
• Denmark: Correlation unchanged at 0.973; RMSE 5.40% vs 5.45% (similar); MAE 3.75%→3.71%; step-change σ error −11.7%→−6.0%.
• France: Correlation 0.975→0.982; RMSE 3.49%→2.97%; MAE 2.63%→2.25%.
• Sweden: Correlation 0.950→0.975; RMSE 6.10%→4.40%; MAE 4.58%→3.21%.
• BPA: Correlation 0.756→0.945; RMSE 18.4%→9.10%; MAE 13.3%→6.06%; step-change σ error −10.4%→+1.5%. ERA5 also yielded smaller mismatches in the standard deviation of one-hour step changes and lower RMSE of duration curves in most cases. Systems with small equivalent system radii and complex terrain (e.g., BPA, req ≈ 91 km) benefited most from ERA5’s higher resolution; flatter, less complex systems like Denmark (req ≈ 166 km) showed smaller gains.

Equivalent system radius (req) examples (km): Germany 329; Denmark 166; France 376; Sweden 471; BPA 91. Country capacity factors over analysed periods: DE 0.186; DK 0.303; FR 0.215; SE 0.329; BPA 0.266.

Individual turbines (1051 Swedish WTs, hourly): ERA5 improved all metrics relative to MERRA-2. Means (medians): correlation 0.868 (0.874) vs 0.802 (0.812); RMSE 15.8% (15.3%) vs 19.1% (18.3%); MAE 10.6% (10.3%) vs 13.2% (12.6%); duration-curve RMSE 1.92% (1.66%) vs 2.94% (2.71%). Monthly generation correlations improved from 0.922 to 0.964 (means). Variability at single sites is underrepresented by both reanalyses: standard deviation of one-hour step changes underestimated by ~49% (MERRA-2) and ~43% (ERA5), as confirmed by power spectral density comparisons showing deficits at sub-daily frequencies.

Long-term correction (893 WTs): Using one baseline year to predict energy in the remaining 1–5 years, errors were substantially reduced by LTC relative to a naïve baseline-CF estimator. Naïve: MAE 7.53%, RMSE 9.25%. MERRA-2 LTC: MAE 3.00%, RMSE 3.89%. ERA5 LTC: MAE 2.37%, RMSE 3.16% (~20% lower than MERRA-2). Using two baseline years lowers errors for both reanalyses by ~16%, implying that one year of measurements LTC’d with ERA5 is more accurate than two years LTC’d with MERRA-2.

The study set out to test whether ERA5 improves wind power modelling relative to MERRA-2 at both aggregated and individual scales. Results confirm that ERA5 yields higher correlations, lower RMSE/MAE, and closer matching of variability metrics across diverse systems. Gains are largest in complex terrain and in smaller, more concentrated systems where higher spatial resolution better captures spatial gradients and temporal dynamics. While individual-turbine variability remains underrepresented due to coarse model resolution, aggregation reduces this mismatch; ERA5 nonetheless narrows the gap in step-change variability compared to MERRA-2. The proposed equivalent system radius offers an interpretable measure linking spatial dispersion to aggregate variability and modelling difficulty, aiding system comparisons and contextualizing performance differences across regions. For industry use cases like asset valuation and performance assessment, ERA5-based LTC reduces uncertainty materially, which can translate into lower financial risk and higher project value. Caution is warranted when comparing across studies due to differing periods and datasets, and because the modelling here was intentionally simple and generic; more tailored models may yield further improvements, potentially more so for MERRA-2 due to larger biases it exhibits. Nevertheless, the magnitude and consistency of ERA5’s advantage suggest practical benefits for both research and operational analyses.

ERA5 outperforms MERRA-2 for wind power modelling at both country and individual-turbine scales. On average across five systems, ERA5 reduced hourly MAE by ~24% and RMSE by ~22; in the BPA area, errors were roughly halved and correlations markedly improved. For 1051 Swedish turbines, ERA5 increased hourly correlations by about 6–7 percentage points and lowered MAE by ~20%. In LTC applications, ERA5 reduced energy prediction errors by ~20% relative to MERRA-2; one year of measurements corrected with ERA5 performed better than two years corrected with MERRA-2. ERA5’s extended temporal coverage (to be 1950–present) further enhances its utility. Future work should pursue: (i) analytic solutions or improved approximations for the equivalent system radius integral; (ii) enriched modelling that incorporates stochastic downtime and multivariate noise to better reproduce spectra, correlations, and maxima for future scenarios (e.g., higher hub heights, lower specific ratings, more offshore capacity); and (iii) exploration of bias corrections and terrain-aware adjustments to further close the gap in single-site variability representation.

• Reanalysis fields cannot resolve local site effects and under-represent single-turbine variability, leading to underestimation of high-frequency fluctuations.
• Hub heights and rotor-specific parameters were unavailable; a single generic power curve (360 W/m²) and constant shear assumptions were used.
• Internal/external losses were modelled with simple fixed adjustments; country-level maxima were enforced using measured maxima, limiting direct applicability to future scenarios without additional stochastic modelling.
• Coordinates for Swedish turbines were estimated via best-fit reanalysis grid points for most units; validation suggests minimal bias, but residual location uncertainty remains.
• The modelling intentionally avoided country-specific tuning; more detailed, tailored models might improve performance (possibly more for MERRA-2).
• Comparisons to other studies are limited by differing time periods and datasets.
• The equivalent system radius formulation requires numerical solution and depends on an assumed exponential correlation decay and chosen length scale (D = 500 km).