EPT-2 Outperforms ECMWF HRES on Solar Radiation in Europe

Written by: Olivier Lam, Physical AI Team, Jua.ai AG

Key Takeaways for European Solar Forecast Buyers

• EPT-2 leads independent benchmarks on surface solar radiation accuracy across the full 0–240 hour forecast range, outperforming ECMWF HRES and other models without post-processing.

• European energy traders can reduce annual imbalance costs by millions; a 1 GW solar portfolio gaining four percentage points of forecast accuracy saves approximately €3 M per year.

• Jua for Energy delivers native solar, wind, and renewables power forecasts for DE, GB, FR, NL, and BE with 15-minute actual-generation refresh and ensemble outputs that beat the ECMWF ENS mean.

• The Athena AI agent converts natural-language queries into briefings, benchmarks, and backtests in under 90 seconds, replacing hours of manual data prep for trading desks.

• EPT-2’s accuracy and refresh cadence translate into lower imbalance costs and faster decisions for European traders managing multi-GW portfolios.

Key Evaluation Criteria for Solar Forecast Software in Europe

European energy traders, meteorologists, and quant developers evaluate solar forecast software on five dimensions. Each dimension maps directly to a cost or workflow risk.

Accuracy on surface solar radiation. SSRD skill at day-ahead and intraday horizons determines imbalance exposure. Higher skill reduces imbalance cost. A 2026 meta-analysis of 5,277 observations across 250 studies found that ensemble-hybrid models deliver the most robust performance, increasing skill scores by 7–27 percentage points relative to time-series baselines, and that day-ahead solar forecasts achieve the largest gains when numerical weather prediction data are used as inputs. These findings explain why platforms that expose ensemble outputs alongside deterministic runs give traders a probabilistic view of generation risk that single-model tools cannot provide.

Update frequency. The European power market trades intraday continuously. A platform that refreshes four times per day leaves traders exposed between runs. EPT-2e updates four times per day. Actual-generation power forecasts on Jua for Energy refresh every 15 minutes, which aligns with intraday trading cadence.

Workflow integration. Raw model outputs in grib format require an in-house pipeline before they become tradeable signals. Platforms that expose a REST API with Apache Arrow support and a Python SDK cut integration time from a quarter to days and reduce ongoing maintenance work.

Spatial resolution. Hub-height and panel-tilt corrections require fine spatial granularity. EPT-2 HRRR produces forecasts at native resolution up to 5 km over Europe. The Jua for Energy product surface supports up to 1 km resolution for comparison use cases and localised analysis.

Europe coverage. Day-ahead and intraday positions in DE, GB, FR, NL, and BE require native power forecasts calibrated to each market’s installed capacity mix. Generic weather variables re-labelled as generation signals do not provide that level of fidelity.

EPT-2 meets these criteria while improving accuracy and refresh speed on the variables that drive imbalance costs.

Head-to-Head Accuracy: EPT-2 vs ECMWF and AI Peers

The benchmark published in arXiv:2507.09703 evaluates deterministic forecast accuracy across 25 models on RMSE against more than 10,000 real ground stations, with no post-processing or station fine-tuning. EPT-2 outperforms ECMWF HRES on surface solar radiation across the full 0–240 hour lead-time range. Microsoft Aurora has no SSRD output, so EPT-2 wins that comparison by default. EPT-2e, the ensemble variant, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time.

The following table shows how EPT-2 stacks up against ECMWF HRES and AI peers across the dimensions that determine trading-desk value: SSRD accuracy, update frequency, Europe power coverage, and ensemble availability.

European energy traders are increasingly using AI tools to anticipate shifts in the ECMWF two-week outlook, the definitive reference for repricing risk around renewable output and system tightness. A platform that outperforms ECMWF HRES on SSRD at every lead time, and delivers that signal hours before the next traditional run, changes the information asymmetry on the trading desk.

Run the benchmark yourself on your region and variable to see how EPT-2 compares with your current provider.

15-Minute Solar Power Forecasts Across Five EU Countries

Most solar forecast platforms deliver generic weather variables re-labelled as generation signals. Jua for Energy takes a different approach and delivers native power forecasts capacity-weighted to each market. These forecasts cover Germany, Great Britain, France, the Netherlands, and Belgium and align directly with installed capacity.

Jua for Energy provides native solar, wind onshore, wind offshore, total wind, total renewables, load, and residual load forecasts across these five markets. Two complementary models run on the same surface. The Fundamental Model combines EPT-2 weather output with installed-capacity data and runs out to 20 days. The Actual Generation Model refreshes every 15 minutes with a 48-hour horizon and lower near-term error.

This surface is not a generic weather variable re-labelled as a generation signal. Forecasts are capacity-weighted via Market Aggregates 2.0, with full cross-model comparison, deltas, disagreements, and heatmaps. The cost impact is measurable. A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves approximately €1.5 M per year. For solar, the savings are even larger, with a 1 GW portfolio gaining four percentage points of accuracy saving approximately €3 M per year, a figure that scales linearly across multi-GW portfolios.

The 15-minute actual-generation refresh means traders see the next signal hours before the next traditional NWP run lands. Traders act before the market does and capture value from earlier information. See the 15-minute refresh in action on a live portfolio walkthrough.

Natural-Language Briefings with the Athena Agent

Athena turns Jua for Energy’s forecast surface into trader-ready briefings. A trader types a question in natural language, such as “what is the solar generation forecast spread across models for southern France tomorrow?” Athena then plans, calls tools, evaluates intermediate outputs, and returns a briefing, a benchmark, a backtest, or a custom widget.

Athena turns raw physics predictions from EPT-2 into actionable outputs by reading market context and modelling participant behaviour. A typical query resolves in approximately 90 seconds. A backtest resolves in approximately 5 minutes. Trading houses and quant desks describe Athena as “another headcount, for free” because it absorbs routine analysis work.

The 7–9 a.m. manual prep routine of downloading grib files, waiting for the meteorologist’s briefing, and stitching together terminal screens compresses into a single workspace open before the market does. See how Athena reshapes your morning prep with a guided session.

API and SDK Integration for Quant Teams

Athena serves traders who need fast answers in natural language. Quant teams building systematic strategies access the same forecast surface programmatically through the API and SDK.

Jua for Energy exposes more than 25 models through a REST API (POST /v1/forecast/data and related endpoints) with Apache Arrow support for large payloads. The Python SDK installs via pip install jua from PyPI and provides forecast access, hindcast and backtesting, and weather-parameter standardisation across all models on the platform.

ENTSO-E grid-data integration is available for European power-market data, covering actual generation, capacity, and PSR classifications across DE, GB, FR, NL, and BE. Hindcast data is available across multiple Jua and third-party models for backtesting systematic strategies. The integration that takes a quant team a quarter to build elsewhere stands up in days on Jua.

Documentation is at docs.jua.ai. The developer dashboard is at developer.jua.ai. Explore the API and SDK on a technical deep-dive tailored to your stack.

When to Choose Jua for Energy vs Point-Solution Vendors

Point-solution SaaS vendors in the solar forecasting space resell processed NWP outputs. These vendors do not own a forecasting model, do not run cross-vendor benchmarks, and do not expose an ensemble or a natural-language agent layer. They fit asset operators who need irradiance data for a single site and have no requirement for multi-model comparison, intraday briefings, or programmatic access to hindcasts.

Jua for Energy fits trading and analytics teams with broader requirements. Typical needs include head-to-head accuracy benchmarking across more than 25 models on the user’s own region and variable, native solar power forecasts in DE, GB, FR, NL, or BE with 15-minute actual-generation refresh, and ensemble probabilistic skill that beats the ECMWF ENS mean. Many teams also require a natural-language agent that turns questions into briefings and backtests in under 90 seconds, plus programmatic API and SDK access for systematic strategies.

Jua serves major utilities across four continents, including some of Europe’s largest energy companies, as well as commodity traders and hedge funds, with sales cycles compressed to as little as two weeks. Review these reference deployments to understand how similar desks use Jua for Energy.

Open-Source vs Enterprise Solar Forecast Options

Free solar forecast tools exist and serve a legitimate purpose. Open-source libraries for irradiance modelling, typically built on clear-sky models, satellite-derived climatologies, or publicly available NWP outputs, are appropriate for research, academic benchmarking, and single-site feasibility studies.

Their limitations are consistent. They provide no ensemble outputs, no intraday refresh at trading cadence, no cross-model benchmarking surface, no agent layer, and no native power forecasts calibrated to European market capacity mixes. For teams whose requirement is a tradeable signal with transparent accuracy provenance and workflow integration, open-source tools act as a starting point rather than a production stack.

Conclusion: Why Trading Desks Adopt Jua

Jua is a foundation model and agent company. EPT, the Earth Physics Transformer, is a general physics foundation model. Athena is an AI agent. Jua for Energy is the first applied product built on both and delivers high-accuracy solar forecasts in production, benchmarked independently in arXiv:2507.09703, and used by Axpo, TotalEnergies, Statkraft, EnBW, EDF, and quant funds across five continents.

These users rely on Jua for measurable gains in forecast accuracy, faster intraday refresh, and integrated workflows for traders and quants. See EPT-2 head-to-head against your current forecast provider on your own regions and variables.

Frequently Asked Questions

What makes EPT-2 more accurate than ECMWF HRES on surface solar radiation?

EPT-2 is a spatiotemporal transformer foundation model trained on more than 5 petabytes of observational data from 120+ sources, including geostationary satellites, surface station networks, and ERA5 reanalysis. It learns the governing physics of the atmosphere, such as conservation of mass, momentum, and energy, directly from observational data in a latent representation that is integrated forward in time.

EPT-2 produces forecasts at native arbitrary lead times rather than rolling forward in fixed 6-hour increments as most peer models do. Rolling compounds error, while EPT-2 avoids that rolling process. The benchmark confirms EPT-2’s lead at every lead time, validated against more than 10,000 real ground stations on open-source StationBench with no post-processing or station fine-tuning. The methodology and results are published in the peer-reviewed technical report arXiv:2507.09703.

How does Jua for Energy differ from Solcast or other solar-specific forecasting vendors?

Solar-specific point-solution vendors process and resell NWP outputs. These vendors do not own an underlying forecasting model, do not publish head-to-head accuracy benchmarks against ECMWF on SSRD at 0–240 hour lead times, and do not expose ensemble probabilistic outputs or a natural-language agent layer.

Jua for Energy is built on EPT-2, a general physics foundation model that Jua owns and operates, with EPT-2e delivering ensemble forecasts that beat the 50-member ECMWF ENS mean on RMSE and CRPS at virtually every lead time. The platform also exposes more than 25 third-party models, including ECMWF HRES, Microsoft Aurora, and GFS GraphCast, under a unified schema, so traders can benchmark Jua against any alternative in seconds rather than managing separate vendor contracts.

Native solar power forecasts in Germany, Great Britain, France, the Netherlands, and Belgium refresh every 15 minutes on actual generation, a cadence that point-solution vendors do not match.

Can Jua for Energy integrate with our existing ECMWF subscription and internal trading pipelines?

Jua for Energy is designed to run alongside existing ECMWF subscriptions, not replace them. ECMWF HRES, ENS, AIFS, and EC46 all run natively on the Jua platform under the same schema as EPT-2, so traders see every model on one screen with one API.

The platform exposes a REST API with Apache Arrow payload support and a Python SDK installable via pip install jua. ENTSO-E grid-data integration covers actual generation, capacity, and PSR classifications across the five European markets Jua for Energy serves. Hindcast data is available across Jua and third-party models for backtesting.

Quant teams pipe Jua forecasts directly into their own systematic models. Utilities and trading houses connect Jua to existing dispatch, risk, and trading tools. Integration that takes a quarter to build elsewhere typically stands up in days.

What is the Athena agent and how does it support solar trading workflows?

Athena is Jua’s AI agent, currently instrumented with the Jua for Energy tool surface. It accepts natural-language objectives, plans a sequence of tool calls, evaluates intermediate outputs, and returns a deliverable such as a briefing, a benchmark, a backtest, or a custom widget.

A trader can ask “what is the solar generation forecast disagreement across models for Belgium this weekend?” and receive an analyst-grade answer in approximately 90 seconds. Backtests against years of historical forecasts run in approximately 5 minutes. Athena auto-generates personalised dashboards on request and removes the manual assembly step that currently consumes the 7–9 a.m. prep window.

Internal meteorology teams freed from daily briefing production report redirecting that capacity to deeper forecast research. Athena is not a narrow weather assistant or a single-purpose energy analyst. It is a domain-agnostic AI agent instrumented for the energy trading tool surface in the current product.

How quickly can we validate EPT-2 accuracy against our current solar forecast provider?

The live benchmark on the Jua platform returns a head-to-head accuracy comparison in seconds from first selection. A prospect picks a region and a variable, such as surface solar radiation, and a time window, selects their current provider alongside EPT-2, and the platform returns the comparison immediately.

No data export, no spreadsheet, and no waiting for a vendor to prepare a slide deck are required. For teams that want a deeper validation, Athena runs a full backtest against years of historical forecasts in approximately 5 minutes. This benchmark moment consistently triggers deals because meteorologists who were sceptical of vendor accuracy claims become internal champions once they run the numbers themselves.