High Resolution Hourly Forecast: EPT2-HRRR vs NOAA HRRR

Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: June 27, 2026

Key Takeaways for Energy Desks

• High-resolution hourly forecasts now sit at the core of energy trading. AI-native models like EPT2-HRRR refresh up to 24 times per day, while traditional NWP remains constrained by HPC cycles.
• EPT2-HRRR delivers ~5 km resolution over Europe with native 100 m hub-height wind and SSRD outputs, avoiding the post-processing that standard NWP models require for energy-specific variables.
• The model runs at ~0.25 kWh inference cost on a single GPU and supports up to 24 daily updates, which keeps rapid-refresh economically viable compared with traditional NWP’s supercomputer requirements.
• EPT-2 outperforms ECMWF HRES on key energy variables across all lead times, with benchmarks documented in arXiv:2507.09703 and independently verifiable through Jua’s live benchmarking platform.
• Start a live benchmark to compare EPT2-HRRR against your current forecast provider on your own region and variables.

Why Resolution and Variables Matter for Energy Trading

Spatial resolution in weather forecasting determines whether a model can resolve the terrain features, coastlines, and land-surface heterogeneity that drive the variables energy traders actually care about. A 25 km grid cell averages over an offshore wind farm, a mountain ridge, and a valley floor simultaneously, which produces a forecast that is accurate for none of them. Operational NWP and AI weather models predominantly output standard meteorological variables such as 2 m temperature and 10 m wind speed rather than energy-specific variables such as hub-height wind speed or solar irradiance, and that gap directly costs energy traders money.

For wind energy, the relevant variable is wind speed at hub height, typically 80 m to 150 m above ground, not the 10 m surface wind that most public NWP outputs report. For solar, the relevant variable is surface solar radiation downwards (SSRD), not a generic cloud-cover proxy. Both use cases require a model with sufficient spatial resolution to capture local orographic and surface effects, and sufficient vertical resolution to distinguish hub-height dynamics from the surface layer.

EPT-2, the flagship model underlying Jua for Energy, natively forecasts at up to 5 km resolution. The Jua for Energy product reaches up to 1 km resolution for operational deployments. ECMWF HRES operates at 9 km, and most AI weather peers publish at roughly 25 km. The resolution gap is not cosmetic; it determines whether a forecast is usable for asset-level dispatch and intraday trading decisions.

See 5 km resolution on your assets by running a live benchmark on your region and variable.

EPT2-HRRR vs NOAA HRRR: Specs That Matter for Traders

The table below compares EPT2-HRRR and NOAA HRRR across four attributes that drive energy trading operations. Every figure is cited inline.

The benchmark anchoring EPT2-HRRR’s accuracy claims is documented in arXiv:2507.09703: EPT-2 outperforms ECMWF HRES on 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation across the full 0–240 hour lead-time range. The ensemble variant, EPT-2e, extends this performance advantage by beating the 50-member ECMWF ENS mean on both RMSE (root mean square error) and CRPS (continuous ranked probability score) at virtually every lead time. These benchmarks explain why Jua for Energy runs alongside ECMWF rather than replacing it, as serious customers keep their ECMWF subscription and add EPT2-HRRR for the intraday cadence and energy-variable coverage that ECMWF HRES does not natively provide at this refresh rate.

NOAA HRRR covers the contiguous United States, plus an Alaska sector. EPT2-HRRR covers Europe. For European energy traders, NOAA HRRR does not function as an operational alternative; it serves as a design reference point for what rapid-refresh NWP can achieve at 3 km resolution, now matched and exceeded by an AI-native physics model at continental scale.

Compare these specs in a live session and see the benchmark numbers on your own portfolio region.

Hub-Height Wind and Other Energy-Relevant Outputs

Hub-height wind, the wind speed at the rotor center of a utility-scale turbine, typically 80 m to 150 m above ground, is the single most consequential variable for wind power forecasting. Most operational NWP and AI weather models output standard meteorological variables rather than hub-height wind speed or solar irradiance, which leaves energy teams to post-process 10 m wind through logarithmic or power-law extrapolation. That extra step introduces error, particularly in complex terrain and during stable atmospheric conditions.

EPT-2 natively outputs wind at 11 height levels from 10 m to 200 m, including 100 m, the standard proxy for modern hub heights. ECMWF’s open data includes 100 m wind components (100u, 100v) and surface solar radiation downwards (ssrd), but at 0.25-degree resolution, roughly 25 km, and four runs per day, which does not provide enough cadence for intraday trading. EPT-2 delivers hourly global updates, breaking the four-times-daily refresh cycle that every other system still follows, while maintaining the energy-variable coverage that ECMWF provides at lower frequency.

Surface solar radiation downwards (SSRD) is the second critical energy variable. Microsoft Aurora has no SSRD output. NOAA HRRR provides solar radiation outputs, but only over CONUS. EPT-2 produces SSRD natively, benchmarked against ground truth in arXiv:2507.09703, and outperforms ECMWF HRES on this variable across the full forecast horizon. Beyond variable coverage, the frequency at which these forecasts refresh determines their practical value for intraday trading.

Update Frequency, Cost, and Dissemination Time

The economics of traditional NWP set a hard ceiling on update frequency. NOAA HRRR runs hourly over CONUS, which represents an engineering achievement for a convection-allowing NWP system, but the underlying model still requires HPC infrastructure that limits global deployment. A single traditional NWP simulation consumes roughly 8,400 kWh and costs €1,000–€20,000 to run. The European supercomputer produces two full ECMWF HRES runs per day, supplemented by smaller runs for a total of roughly four global forecasts per 24 hours.

EPT-2 family models, including EPT2-HRRR, run inference on a single GPU in minutes at about 0.25 kWh and $0.20–$15 per simulation, which is approximately four orders of magnitude cheaper than traditional NWP at run time. This cost structure makes 24 daily updates economically viable. EPT2-HRRR delivers up to 24 updates per day over Europe at ~5 km resolution. Between those updates, traders on Jua for Energy are not looking at stale numbers from the previous 06 UTC run; they are looking at a forecast that refreshed within the last hour.

Dissemination time compounds the cadence advantage. A typical Jua for Energy run completes approximately 2.5 hours ahead of competing operational runs at the same cycle. In intraday gas and power markets where the trade window opens and closes in minutes, receiving the updated forecast before the market re-prices defines the practical trading edge.

Accessing and Benchmarking High-Resolution Forecasts

Jua for Energy exposes EPT2-HRRR and more than 25 additional models, including ECMWF HRES, ECMWF ENS, ECMWF AIFS, NOAA GFS, Microsoft Aurora, GFS GraphCast, and DWD ICON, through a unified REST API and Python SDK. The live benchmarking surface on the Jua platform returns a head-to-head accuracy comparison on any region, variable, and time window in seconds.

Quant developers and engineering teams can integrate quickly:

pip install jua

The REST API uses POST /v1/forecast/data with Apache Arrow support for large payloads. Hindcast data is available across multiple Jua and third-party models for backtesting. Documentation is at docs.jua.ai and the developer dashboard at developer.jua.ai. Jua serves major utilities across four continents, including some of Europe’s largest energy companies, as well as commodity traders and hedge funds, with customers such as Axpo, TotalEnergies, Statkraft, EnBW, EDF, and Hydro-Québec.

Athena, Jua’s AI agent instrumented with the Jua for Energy tool surface, turns a natural-language query into a benchmark, backtest, or custom widget in approximately 90 seconds. A full backtest against years of historical forecasts runs in about 5 minutes.

Get API access and run EPT2-HRRR head-to-head against your current forecasts.

Common Evaluation Pitfalls for High-Resolution Sources

Evaluating resolution without checking variable coverage. A 3 km model that outputs only 10 m wind and a generic cloud-cover proxy is less useful for hub-height wind forecasting than a 5 km model that natively outputs 100 m wind and SSRD. Resolution is one dimension, and variable coverage is another, and both require joint evaluation.

Ignoring dissemination time. Two models with identical spatial resolution and update frequency can differ by hours in when their output reaches the trader. A forecast that arrives after the intraday auction has closed has zero trading value regardless of its accuracy. Dissemination time, not just run frequency, determines operational utility.

Accepting vendor-provided accuracy graphics without running independent benchmarks. Meteorologists who have evaluated Jua for Energy describe the live benchmark moment as the deal trigger. They run the comparison themselves on their own region and variable, and the numbers speak. Any evaluation that relies solely on vendor-supplied charts without an independent head-to-head test on operationally relevant variables remains incomplete.

Omitting hindcasts from the evaluation. A model that cannot supply years of historical forecast data cannot be backtested. Quant teams that skip hindcast availability in their evaluation criteria discover the gap only after procurement, when they attempt to validate a systematic strategy and find that the data does not exist.

Treating geographic coverage as binary. NOAA HRRR covers CONUS. European energy traders evaluating rapid-refresh options cannot substitute NOAA HRRR for a European-coverage equivalent. EPT2-HRRR covers Europe at ~5 km resolution with up to 24 daily updates, which forms the operationally relevant comparison for European power and gas markets.

FAQ

What is EPT2-HRRR and how does it differ from NOAA HRRR?

EPT2-HRRR is Jua’s high-resolution rapid-refresh model variant, operating at approximately 5 km spatial resolution over Europe with up to 24 updates per day. NOAA HRRR is NOAA’s High-Resolution Rapid Refresh NWP system, operating at 3 km resolution over CONUS, plus an Alaska sector, with hourly runs where the 00/06/12/18 UTC cycles receive longer forecasts. The key operational differences are geographic coverage, Europe versus CONUS, update cadence, up to 24 versus hourly, inference cost, about 0.25 kWh per simulation on a single GPU for EPT-2 family models versus HPC infrastructure, and energy-variable coverage, 100 m wind and SSRD natively versus standard meteorological outputs that require post-processing.

Does Jua for Energy replace ECMWF?

No. Jua for Energy runs alongside ECMWF, not in place of it. Most serious customers keep their ECMWF subscription and add Jua for Energy for the intraday cadence, hub-height wind coverage, and energy-variable specificity that ECMWF HRES does not provide at rapid-refresh frequency. ECMWF AIFS, ECMWF’s own AI model, runs on the Jua platform as one of more than 25 models available through a unified schema. Jua for Energy displaces the plumbing around the ECMWF feed, including the in-house grib pipeline, the manual benchmarking, the morning-briefing assembly, and the spreadsheet stitching.

How is EPT-2 accuracy validated?

EPT-2 is benchmarked against more than 10,000 real ground stations using open-source StationBench, with no post-processing or station fine-tuning. Results are published in a peer-reviewed technical report at arXiv:2507.09703 and show consistent outperformance of ECMWF HRES across all energy-relevant variables and lead times. EPT-2e, the ensemble variant, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time. The live benchmarking surface on the Jua platform allows any prospect to replicate a head-to-head comparison on their own region and variable in seconds.

What energy-specific variables does Jua for Energy provide that standard NWP sources do not?

Jua for Energy provides wind at 11 height levels from 10 m to 200 m, including 100 m hub-height wind, and surface solar radiation downwards (SSRD) natively, without requiring post-processing extrapolation from 10 m surface wind. Most public NWP sources output standard meteorological variables and require energy teams to apply logarithmic or power-law extrapolation to estimate hub-height wind, which introduces additional error. Microsoft Aurora has no SSRD output at all. Jua for Energy also provides power forecasts for solar, wind onshore, wind offshore, total wind, total renewables, load, and residual load across five European countries, refreshing every 15 minutes for actual generation.

How do quant developers integrate Jua for Energy into existing pipelines?

Quant developers integrate Jua for Energy by running pip install jua to install the Python SDK from PyPI. The REST API exposes more than 25 models through a single schema with Apache Arrow support for large payloads. Hindcast data is available across multiple Jua and third-party models for backtesting systematic strategies. The developer dashboard is at developer.jua.ai and full documentation at docs.jua.ai. Quant teams pipe Jua forecasts directly into their own trading and risk systems, and integration that takes a quarter to build elsewhere typically stands up in days. ENTSO-E grid data is available via direct integration for European power-market context.

Conclusion: Turning High-Resolution Forecasts into P&L

High resolution hourly forecast sources differ across four dimensions that determine their operational value for energy trading: spatial resolution, update cadence, energy-variable coverage, and integration fit. NOAA HRRR established the design standard for rapid-refresh NWP at 3 km with hourly runs over CONUS, plus an Alaska sector. EPT2-HRRR extends that standard to Europe at ~5 km resolution with up to 24 daily updates, native 100 m wind and SSRD coverage, and inference economics that keep the refresh frequency sustainable without HPC infrastructure.

The benchmark is concrete and independently verifiable. The study documented in arXiv:2507.09703 demonstrates consistent outperformance of ECMWF HRES on every variable that drives an energy P&L across the full forecast horizon. A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves roughly €1.5 M per year under typical hedging and penalty structures, and that economics scales linearly across multi-GW portfolios.

Jua is a foundation model and agent company, and Jua for Energy is the first applied product. The architecture learns physics, and the domain is a variable. Meteorologists, energy traders, and quant developers who need a production-grade high-resolution hourly forecast that runs alongside ECMWF and integrates directly into existing pipelines start the evaluation with a live benchmark.

Begin with a benchmark and run EPT2-HRRR head-to-head against your current forecast provider on your own region and variables.