Written by: Olivier Lam, Physical AI Team, Jua.ai AG
Key Takeaways for European Energy Traders
- ECMWF HRES is the deterministic flagship model at ~9 km horizontal resolution with 137 vertical levels out to 10 days, and it remains the institutional baseline for European energy traders.
- Accuracy over Europe degrades most sharply between 72 h and 120 h for wind, temperature, and solar radiation, with effective resolution dropping further over complex terrain such as the Alps.
- Independent StationBench verification shows Jua’s EPT-2 outperforming HRES on 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation at every lead time from 0–240 h.
- Energy traders can keep HRES for its validation record while running Jua for Energy alongside it to replace manual pipelines, benchmarking, and dashboard stitching.
- Book a demo with Jua to benchmark your own region and variables against 25+ models in under five minutes.
HRES Grid and Vertical Structure over Europe
HRES operates on a cubic octahedral reduced Gaussian grid at approximately 9 km horizontal spacing, with 137 vertical levels extending from the surface to 0.01 hPa (approximately 80 km altitude). The table below summarises the operational specification and highlights the update frequency and forecast horizon that shape how traders use HRES for intraday versus day-ahead positioning.
| Parameter | ECMWF HRES Specification | Notes | Source |
|---|---|---|---|
| Horizontal grid spacing | ~9 km (TCo1279) | Cubic octahedral reduced Gaussian grid | ECMWF IFS Documentation |
| Vertical levels | 137 model levels | Surface to 0.01 hPa (~80 km) | ECMWF IFS Documentation |
| Forecast horizon | 10 days (240 h) deterministic | Extended range via ENS | ECMWF IFS Documentation |
| Update frequency | 2–4 runs per day | Full-resolution runs at 00Z and 12Z | ECMWF IFS Documentation |
The 137-level vertical structure gives HRES fine resolution in the planetary boundary layer, which is critical for surface wind and temperature, and in the upper troposphere, where jet-stream dynamics govern medium-range predictability. Each full HRES simulation consumes approximately 8,400 kWh of compute and costs between €1,000 and €20,000 to run on high-performance computing infrastructure, which constrains operational refresh to two to four cycles per day.
Europe Forecast Accuracy by Lead Time (0–240 h)
HRES accuracy over Europe degrades with lead time in a consistent pattern across four variables critical to energy trading: 10 m wind speed, 100 m wind speed (hub-height for most onshore turbines), 2 m temperature, and surface solar radiation downwards (SSRD). The verification below is drawn from EPT-2 (arXiv:2507.09703), evaluated against more than 10,000 real ground stations using open-source StationBench methodology with no post-processing or station fine-tuning, the same methodology that benchmarks EPT-2 against HRES across the full 0–240 h range. The table shows that EPT-2 maintains an accuracy advantage over HRES across all four variables and all lead times, with no crossover point where HRES becomes superior.
| Variable | Lead Time | HRES RMSE (indicative) | EPT-2 RMSE vs HRES | Source |
|---|---|---|---|---|
| 10 m wind speed | 24 h | Reference | EPT-2 outperforms HRES | arXiv:2507.09703 |
| 10 m wind speed | 72 h | Reference | EPT-2 outperforms HRES | arXiv:2507.09703 |
| 10 m wind speed | 120 h | Reference | EPT-2 outperforms HRES | arXiv:2507.09703 |
| 10 m wind speed | 240 h | Reference | EPT-2 outperforms HRES | arXiv:2507.09703 |
| 100 m wind speed | 24–240 h | Reference | EPT-2 outperforms HRES at every lead time | arXiv:2507.09703 |
| 2 m temperature | 24–240 h | Reference | EPT-2 outperforms HRES at every lead time | arXiv:2507.09703 |
| Surface solar radiation | 24–240 h | Reference | EPT-2 outperforms HRES at every lead time | arXiv:2507.09703 |
Accuracy decay is not linear. Wind skill degrades most sharply between 72 h and 120 h, where synoptic-scale predictability limits begin to bind. Temperature skill holds longer over flat terrain but deteriorates faster over complex orography. SSRD errors compound with cloud-cover uncertainty, which grows non-linearly beyond day 4.
Terrain Effects on HRES Effective Resolution
Nominal 9 km grid spacing does not translate uniformly into 9 km effective resolution across Europe’s terrain. Over the Alps and Pyrenees, the effective resolution of HRES degrades substantially relative to flat-terrain performance. Three mechanisms drive this degradation.
First, orographic parameterisation. At HRES’s 9 km spacing, individual Alpine valleys and ridge lines are sub-grid features. The model represents them through parameterisation schemes rather than explicit dynamics, which introduces systematic errors in valley-wind channelling, cold-air pooling, and foehn events.
Second, representativeness error. Ground-truth stations in complex terrain sit at elevations that differ from the model’s smoothed orography by hundreds of metres, which inflates apparent RMSE even when the model’s free-atmosphere dynamics are correct.
Third, precipitation phase and timing. Convective initiation over the Alps and Pyrenees is sensitive to boundary-layer convergence at scales below 9 km, which causes timing errors in precipitation onset that propagate into solar radiation and wind forecasts downstream.
Over flat terrain such as the North German Plain, the Paris Basin, and the Po Valley, HRES performs closer to its nominal specification. The practical implication for energy traders is clear. A wind portfolio in northern Germany will see better HRES skill than an equivalent portfolio in the Swiss or Austrian Alps, and the accuracy gap between HRES and higher-resolution models widens as terrain complexity increases. EPT-2 HRRR, Jua’s high-resolution rapid-refresh variant, operates at approximately 5 km native resolution over Europe and reduces the representativeness gap in complex terrain.
Head-to-Head Comparison: HRES, GFS, ICON-EU, and EPT-2
The terrain effects described above apply differently across the four models most commonly evaluated by European energy traders, and resolution alone does not determine effective skill, although it sets the ceiling over complex orography. The table below compares these four models on the dimensions that matter for operational energy forecasting. The comparison shows that while ICON-EU matches HRES on some variables within its European domain, only EPT-2 consistently outperforms HRES across all four energy-critical variables at every lead time.
| Capability | ECMWF HRES | NOAA GFS | DWD ICON-EU | EPT-2 (Jua for Energy) |
|---|---|---|---|---|
| Horizontal resolution | ~9 km | ~13 km | ~6.5 km (Europe domain) | ~5 km (EPT-2 HRRR, Europe) |
| Vertical levels | 137 | 127 | 60 | Learned latent representation |
| 10 m wind accuracy vs HRES (0–240 h) | Reference | Below HRES | Comparable to HRES over Europe domain | Outperforms HRES at every lead time |
| 100 m wind accuracy vs HRES (0–240 h) | Reference | Below HRES | Limited hub-height output | Outperforms HRES at every lead time |
| 2 m temperature accuracy vs HRES (0–240 h) | Reference | Below HRES | Comparable to HRES over Europe domain | Outperforms HRES at every lead time |
| Surface solar radiation (SSRD) | Reference | Available | Available | Outperforms HRES at every lead time |
| Update frequency | 2–4×/day | 4×/day | 4×/day (EU domain) | Up to 24×/day (EPT-2 RR) |
| Forecast horizon | 10 days (HRES); 15 days (ENS) | 16 days | 5 days (EU domain) | 20 days deterministic; 60 days ensemble |
| Inference cost per run | ~€1,000–€20,000 on HPC | Publicly funded | Publicly funded | ~$0.20–$15 on a single GPU |
Jua’s EPT-2 outperforms leading AI weather models and traditional numerical baselines across all forecast horizons on RMSE. Serious customers, including Axpo, TotalEnergies, Statkraft, EnBW, EDF, and Hydro-Québec, keep their ECMWF subscription and run Jua for Energy alongside it. Jua for Energy does not replace ECMWF. It displaces the plumbing around it: the in-house grib pipeline, the manual benchmarking, the morning-briefing analyst, and the dashboard stitching.
When Energy Traders Should Switch to Ensembles
Beyond approximately 72–96 h, deterministic skill from HRES or any other single-trajectory model degrades to the point where probabilistic guidance becomes operationally superior. The ECMWF Ensemble (ENS) runs 50 perturbed members and is the gold standard for probabilistic NWP out to 15 days. ENS mean RMSE and CRPS are the reference metrics for ensemble verification.
EPT-2e, the ensemble variant of Jua’s EPT-2, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time, with 10 published members and a 60-day horizon. That ensemble advantage becomes operationally critical when traders position around multi-day wind ramps, cold snaps, or solar droughts, events where tail-risk quantification matters more than point accuracy. The practical decision rule follows directly. Use deterministic HRES or EPT-2 for intraday and day-ahead precision, where you need the sharpest single-trajectory forecast. Then switch to ENS or EPT-2e for day 4 and beyond, where spread quantification captures the range of possible outcomes that deterministic models miss.
Frequently Asked Questions
What is the horizontal grid spacing of ECMWF HRES?
ECMWF HRES operates on a cubic octahedral reduced Gaussian grid at approximately 9 km horizontal spacing (TCo1279 spectral truncation). This is the nominal resolution, and effective resolution over complex terrain such as the Alps and Pyrenees is lower because sub-grid orographic features are represented through parameterisation rather than explicit dynamics. For comparison, EPT-2 HRRR, Jua’s high-resolution rapid-refresh variant, operates at approximately 5 km native resolution over Europe.
How many vertical levels does ECMWF HRES have?
HRES uses 137 model levels, extending from the surface to approximately 0.01 hPa (around 80 km altitude). The level spacing is finest in the planetary boundary layer, where surface wind, temperature, and moisture are most directly relevant to energy trading, and coarser in the stratosphere. The 137-level configuration has been in place since the 2013 model cycle upgrade and remains the operational standard as of 2026.
How accurate is ECMWF HRES for wind forecasting over Europe?
HRES is the benchmark for European wind forecasting and has held that position for four decades. Skill degrades with lead time, and 10 m and 100 m wind RMSE increases most sharply between 72 h and 120 h, where synoptic-scale predictability limits bind. Over flat terrain such as the North German Plain and the Paris Basin, HRES performs close to its nominal specification. Over complex terrain, effective skill is lower. EPT-2, evaluated against more than 10,000 ground stations using StationBench with no post-processing, outperforms HRES on 10 m wind and 100 m wind across all lead times, as documented earlier. A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves approximately €1.5 million per year under typical hedging and imbalance-cost structures.
Does ECMWF HRES cover surface solar radiation for energy trading?
Yes. HRES produces surface solar radiation downwards (SSRD) as a standard output variable, and traders and asset operators use it widely. SSRD skill degrades with cloud-cover uncertainty, which grows non-linearly beyond day 4. EPT-2 outperforms HRES on SSRD at every lead time from 0 to 240 h, as documented in arXiv:2507.09703. Microsoft Aurora, for reference, produces no SSRD output at all, so EPT-2 wins that comparison by default.
Should energy traders replace ECMWF HRES with an AI model?
No. The operationally sound approach is to run both. ECMWF HRES carries forty years of institutional validation, a 137-level vertical structure, and a global verification record that no AI model has yet matched in breadth of operational deployment. Jua for Energy is designed to run alongside HRES, not to replace it. What Jua for Energy displaces is the infrastructure around the HRES feed: the grib-file pipeline, the manual benchmarking, the morning-briefing analyst, and the dashboard stitching. ECMWF AIFS, ECMWF’s own AI model, runs natively on the Jua platform and is available in the same workspace as EPT-2 and HRES. The comparison is built in, and the decision belongs to the trader.
Conclusion: How to Use HRES and EPT-2 Together
ECMWF HRES remains the forty-year benchmark, with 9 km horizontal spacing, 137 vertical levels, and a verification record that defines the reference standard for European energy forecasting. Its accuracy decays measurably by lead time, most sharply between 72 h and 120 h on wind, and its effective resolution degrades over complex terrain such as the Alps and Pyrenees, where sub-grid orographic features cannot be resolved explicitly at 9 km.
The evaluation lens for any operational forecasting decision covers three dimensions: model capability, which means accuracy at the lead times that map to your trade horizon; operational usability, which covers update frequency, dissemination latency, and API access; and reliability, which includes peer-reviewed verification, physics-grounded architecture, and institutional track record. HRES scores highest on reliability by institutional history. EPT-2 scores highest on capability, maintaining the accuracy advantage documented in the verification section across all four energy-critical variables, and on operational usability, with up to 24 updates per day versus HRES’s two to four. The accuracy gains documented above translate to the P&L impact described earlier, which is meaningful at portfolio scale.
The next step is not a vendor presentation. It is a benchmark on your own region and your own variables, against your current provider, in less than five minutes.