Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: July 11, 2026
Key Takeaways
- Traditional NWP systems update only 2–4 times daily, which leaves energy traders with stale forecasts and higher imbalance costs.
- Physics-constrained AI models like EPT-2 deliver up to 24 daily updates at 5 km resolution and beat ECMWF HRES on wind, solar, and temperature.
- Lower inference costs (0.25 kWh vs. 8,400 kWh) make rapid-refresh forecasts economically viable and aligned with 15-minute intraday trading.
- The Athena agent turns manual prep into ~90-second natural-language queries, auto-generating briefings and backtests that replace hours of work.
- Book a demo with Jua to run a live benchmark against your current provider in under five minutes and see the accuracy edge firsthand.
The Core Problem: Intraday Trading on Stale Weather Data
A single NWP simulation consumes approximately 8,400 kWh of compute and costs between €1,000 and €20,000 to run on high-performance computing infrastructure. That compute ceiling limits ECMWF and NOAA to two full global runs per day, with smaller supplementary cycles bringing the practical total to roughly four forecasts per 24 hours. NOAA’s GFS and ECMWF typically update forecasts every 12 hours, using supercomputers to integrate fluid dynamics and thermodynamics equations forward in time, which has created a hard physical and economic constraint for the energy industry over four decades.
The consequences for intraday trading are direct. Positions fixed in day-ahead markets are frequently wrong by the time power flows because of weather-dependent renewable volatility, so participants must adjust positions continuously in intraday markets to reduce imbalance costs. This continuous adjustment happens at 15-minute resolution under EU electricity market design reform, with cross-zonal gate closure reduced to 30 minutes, which tracks renewable volatility more closely but demands significantly higher data speed and tighter timing than day-ahead markets. That timing pressure exposes the core problem: a forecast that was current at 06:00 UTC is four to six hours stale by the time the intraday window opens, and any wind ramp or solar dip that emerged in that interval is already priced by whoever saw it first.
The market-sizing economics are concrete. A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves approximately €1.5 million per year in European energy markets. A 1 GW solar portfolio at the same accuracy gain saves approximately €3 million per year. Multi-gigawatt portfolios scale these figures linearly.
Book a demo to see EPT-2 head-to-head against your current forecast provider.
Why Physics-Constrained Foundation Models Break the Compute Barrier
The compute-cost barrier that limits NWP to 2–4 daily runs requires a fundamentally different architecture to overcome it. Physics-constrained foundation models are spatiotemporal transformers trained directly on observational data, including satellite feeds, surface station networks, radar, ocean buoys, and reanalysis archives, in a latent representation that respects conservation laws governing mass, momentum, and energy. This architecture differs categorically from unconstrained large language models, which operate on discrete symbolic tokens and can produce outputs that violate physical laws. It also differs from research-only AI weather outputs such as Microsoft Aurora or Google DeepMind GraphCast, which are published as raw model files without productised refresh schedules, ensembles, or workflow tooling.
The EPT (Earth Physics Transformer) family is Jua’s general physics foundation model. It is domain-agnostic by architecture; what changes from one physical system to the next is the data and the fine-tune. The atmosphere is the first physical system EPT has been fine-tuned for, and energy trading is the first market Athena, Jua’s AI agent, has been instrumented for. The EPT family deployed inside Jua for Energy includes EPT-2 (deterministic flagship, four runs per day, 20-day horizon), EPT-2e (ensemble variant, updated daily, 60-day horizon), EPT-2 RR (rapid refresh, up to 24 runs per day), and EPT-2 HRRR (high-resolution rapid refresh, native 5 km resolution over Europe). Training data spans more than 5 petabytes across 120+ distinct sources, with proprietary station coverage exceeding 10,000 stations in EPT-2.
A 2026 study led by Karlsruhe Institute of Technology and the University of Geneva found that pure data-driven AI models without physics constraints systematically underestimate the intensity and frequency of extreme weather events, with underestimation increasing as events exceed training-data records. EPT addresses this directly. Its outputs are physically constrained by construction, not by post-processing, so the architecture cannot produce outputs that violate the conservation laws governing the real atmosphere.
Update Cadence: EPT-2 RR vs Traditional NWP Cycles
Traditional NWP delivers two to four global forecasts per 24 hours. EPT-2 RR delivers up to 24 updates per day. EPT-2 HRRR delivers the same hourly cadence at the 5 km native resolution established earlier over Europe. EPT-2e is updated daily. Actual-generation power forecasts inside Jua for Energy refresh every 15 minutes with a 48-hour horizon.
The dissemination advantage compounds the update-frequency advantage. A typical Jua for Energy run completes approximately 2.5 hours ahead of competing operational runs at the same cycle. EPT-2 delivers hourly global weather updates at six times higher temporal and spatial resolution than comparable AI models. Customers running Jua for Energy alongside their existing NWP subscriptions receive the next forecast hours before the next traditional run lands, which creates a structural edge in any intraday window where model revisions reprice the market.
Intraday trading on EPEX SPOT grew 22% from 176 TWh in 2023 to 215 TWh in 2024 as renewable variability drives continuous re-optimisation needs. A forecast stack that refreshes 24 times per day aligns structurally with that market, while one that refreshes four times does not.
Performance Benchmarks: EPT-2 vs ECMWF and Aurora
EPT-2 outperforms ECMWF HRES on every lead time across the full 0–240 hour range for 10 m wind speed, 100 m wind speed, 2 m temperature, and surface solar radiation (SSRD). EPT-2e, the ensemble variant, beats the 50-member ECMWF ENS mean on both RMSE (root mean square error) and CRPS (continuous ranked probability score, a probabilistic skill metric) at virtually every lead time. Both results are documented in peer-reviewed technical reports: arXiv:2507.09703 for EPT-2 and arXiv:2410.15076 for EPT-1.5. Evaluation methodology uses open-source StationBench against more than 10,000 real ground stations, with no post-processing or station fine-tuning.
EPT-2 also beats Microsoft Aurora on 10 m wind, 100 m wind, and 2 m temperature across the full 0–240 hour range. Aurora produces no SSRD output at all, which makes EPT-2 the only AI model in production delivering the full variable set that drives a European energy P&L. Spatial resolution reaches 5 km natively for EPT-2 HRRR over Europe, compared to 9 km for ECMWF HRES and approximately 25 km for Aurora at published resolution.
ECMWF’s own AIFS ensemble runs four times daily at approximately 30 km resolution, generates forecasts faster than classical NWP, and consumes a fraction of the energy, yet raw IFS ensemble forecasts still substantially outperformed it on mean CRPS and MAE for 10 m wind speed across all lead times up to 15 days in independent evaluation. EPT-2e outperforms the ECMWF ENS mean on the same probabilistic metrics, which establishes a clear performance hierarchy: EPT-2e above ECMWF ENS, and ECMWF ENS above AIFS on raw probabilistic wind skill.
Book a demo to run a live benchmark on your region and variables in under 5 minutes.
Athena Agent: From Natural-Language Questions to Trading Briefings
Athena is Jua’s AI agent, currently instrumented with the Jua for Energy tool surface. A trader types a question in natural language, such as “what is the 100 m wind forecast spread across models for northern Germany tonight?” or “backtest a wind-ramp strategy on EPT-2e over the last two winters”, and Athena plans, calls tools, evaluates intermediate outputs, and returns the answer, the underlying widget, or the full backtest report. Typical queries resolve in approximately 90 seconds, and backtests in approximately 5 minutes.
Athena turns raw physics predictions from EPT-2 into actionable trading intelligence by reading market context and modelling the implications of forecast shifts. Day-Ahead and Intraday briefings auto-refresh on every new model run, covering model consensus across 25+ models, model delta since the previous run, convergence tracking, and price implications, already written in. The 7–9 a.m. manual prep routine, which includes downloading grib files, processing through brittle in-house pipelines, and waiting for the meteorologist’s briefing, compresses into a single workspace open before the market does.
The live benchmark moment usually triggers the deal for Jua for Energy customers. A prospect selects their region and their current provider, and the Jua platform returns a head-to-head accuracy comparison in seconds. The objection shifts from “is this real?” to “how fast can we sign?” Trading houses and quant desks describe Athena as “another headcount, for free”.
Comparative View: How Jua for Energy Stacks Up
Update frequency. EPT-2 RR delivers up to 24 updates per day. ECMWF HRES maintains its traditional cadence of 2–4 runs per day. Aurora and GraphCast operate on a research cadence of approximately four runs per day with no productised operational schedule. The gap between 4 and 24 daily updates is the gap between a stale intraday view and a current one.
Inference cost. A single EPT-2 inference runs at approximately 0.25 kWh and $0.20–$15 on a single GPU in minutes, which reflects the four-orders-of-magnitude cost advantage over traditional NWP described earlier. That cost asymmetry is what makes 24 daily updates economically viable for EPT-2 and structurally impossible for classical NWP. The same efficiency advantage extends to training: EPT-2 was trained on 8 × H100 GPUs over 10 days, while Microsoft Aurora required 32 × A100 GPUs over 18 days.
Workflow integration. Jua for Energy exposes 25+ models through a REST API with Apache Arrow support for large payloads and a Python SDK installable via pip install jua. Hindcast data is available across multiple Jua and third-party models for backtesting. Aurora and GraphCast deliver raw model files without ensembles, hindcasts, or productised tooling, so the engineering team must build the pipeline. Jua for Energy stands that integration up in days.
Benchmarking transparency. The Jua platform hosts 25+ models, including 10 proprietary AI from the EPT family plus 15 third-party NWP and AI models such as ECMWF HRES, ENS, AIFS, NOAA GFS, DWD ICON, Aurora, and GraphCast, on a single benchmarking surface. Any region, any variable, any time window, head-to-head, in seconds. No AI peer offers an equivalent. EPT-2 and EPT-2e performance claims are documented in peer-reviewed arXiv reports, not vendor graphics.
Ensemble depth. EPT-2e outperforms the ECMWF ENS mean on the same probabilistic metrics referenced earlier. No AI weather peer has shipped a productised ensemble equivalent. Aurora has no ensemble product. GraphCast has no ensemble product.
Native any-Δt forecasting. EPT-2 produces forecasts at arbitrary lead times. Aurora and most AI peers are trained on a fixed 6-hour grid and roll forward in 6-hour steps, which compounds error at each step. EPT-2 does not roll.
Risk, Validation, and Integration Considerations
Data quality and physics validity. The objection that AI weather models hallucinate or violate conservation laws applies to unconstrained architectures, not to EPT. EPT is a spatiotemporal transformer trained on observational physics, and its latent representation is integrated forward in time under conservation-law constraints. The validation is external and reproducible through StationBench against 10,000+ real ground stations, published on arXiv, with no post-processing. The 2026 KIT/University of Geneva study recommends parallel use of physics-informed and classical NWP approaches for high-risk applications, which matches how Jua for Energy is deployed: alongside ECMWF, not instead of it.
Integration with existing pipelines. Jua for Energy does not require replacing the existing ECMWF subscription. ECMWF AIFS runs on the Jua platform. The REST API and Python SDK expose all 25+ models under a unified schema, so existing pipelines that consume ECMWF outputs can route Jua forecasts through the same interface without re-engineering. ENTSO-E grid data integrates directly for European power-market context.
Validation standards for regulated environments. Procurement teams at regulated utilities evaluate the platform against internal risk and regulatory frameworks. The practical evaluation criteria are clear: run the live benchmark on the company’s most stakes-relevant region and variable (approximately 5 minutes to result), request hindcast data for the relevant historical period, and review the arXiv technical reports for methodology. The benchmark provides the proof-of-value, and it runs on the prospect’s own data, not on vendor-selected examples.
FAQ: EPT-2 RR, AIFS, and Integration
What is the difference between EPT-2 RR and standard EPT-2?
EPT-2 is Jua’s deterministic flagship model, running four times per day with a 20-day forecast horizon. EPT-2 RR is the rapid-refresh variant, running up to 24 times per day. EPT-2 HRRR is the high-resolution rapid-refresh variant, delivering the same hourly cadence at native 5 km resolution over Europe. All three are members of the EPT family, which share the same general physics foundation model architecture and differ by run cadence and spatial resolution. EPT-2e is the ensemble variant that is updated daily with a 60-day horizon and beats ECMWF ENS on RMSE and CRPS, as documented earlier.
How does Jua for Energy compare to ECMWF AIFS?
ECMWF AIFS is ECMWF’s own AI-based forecasting system, available on the Jua platform alongside EPT and 23 other models. The AIFS ensemble runs four times daily at approximately 30 km resolution. EPT-2 outperforms ECMWF HRES, the benchmark AIFS is measured against, on every lead time for 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation across 0–240 hours. EPT-2e is the ensemble variant that beats ECMWF ENS on RMSE and CRPS, which reflects the performance advantage documented earlier. EPT-2 RR updates up to 24 times per day versus AIFS’s four. Jua for Energy does not replace ECMWF; it runs alongside the incumbent feed and displaces the plumbing around it.
What integration work is required to connect Jua for Energy to an existing trading system?
Integration requires installing the Python SDK via pip install jua or connecting to the REST API at query.jua.ai/docs. The API exposes all 25+ models through a single schema with Apache Arrow support for large payloads. Hindcast data is available for backtesting across multiple Jua and third-party models. ENTSO-E grid data integrates directly for European power-market context. Quant teams that have built ingestion pipelines for raw Aurora or GraphCast outputs typically stand up the Jua for Energy integration in days rather than the quarter those raw-output pipelines required. Full documentation is at docs.jua.ai.
Who are the typical users of Jua for Energy?
Jua for Energy serves three archetypes: regulated utilities (EDF, EnBW, Statkraft) whose meteorology teams evaluate forecast quality and whose trading desks need automatic briefings and divergence alerts; physical trading houses (TotalEnergies, Axpo) whose quant pods consume API-first forecast data and whose analysts use Athena for natural-language queries; and capital-markets and quantitative funds whose developers pipe Jua forecasts and hindcasts directly into systematic strategies via the Python SDK. Within each archetype, the buying roles include the meteorologist (technical evaluator), the quant developer (pipeline builder), the trader (daily user), and the senior decision-maker (contract signer).
Is Jua for Energy only an atmospheric forecasting product?
Jua for Energy is the first applied product from Jua’s foundation model and agent platform. The underlying architecture, EPT as a general physics foundation model and Athena as an AI agent, is domain-agnostic. The atmosphere is the first physical system EPT has been fine-tuned for, and energy trading is the first market Athena has been instrumented for. The same EPT model that learns atmospheric dynamics already predicts plasma behaviour inside a tokamak. Jua’s roadmap extends to other physical-economy domains, including plasma fusion, aerospace, materials, and fluids, each shipped as a new vertical product on the same horizontal platform. The relationship mirrors the one Anthropic has to Claude Code, a horizontal AI platform with a flagship vertical product.
Conclusion: Checklist for Solving the Stale-Forecast Problem
The stale-forecast problem is structural. Two to four NWP runs per day reflect a compute economics constraint that has held for forty years and cannot be resolved by faster pipelines or better dashboards built on top of the same underlying cycle. Physics-constrained foundation models with agent layers resolve it at the source. EPT-2 RR runs up to 24 times per day at approximately 0.25 kWh per inference and delivers forecasts that outperform ECMWF HRES on every lead time for the variables that drive a European energy P&L.
The evaluation criteria for any buyer considering this category are straightforward:
- Run a live benchmark on your highest-stakes region and variable against your current provider (approximately 5 minutes on the Jua platform).
- Verify ensemble probabilistic skill against ECMWF ENS mean on RMSE and CRPS (EPT-2e beats the 50-member ENS mean at virtually every lead time).
- Confirm hindcast availability for the historical period your backtesting strategy requires.
- Review the peer-reviewed technical reports ( arXiv:2507.09703 for EPT-2 and arXiv:2410.15076 for EPT-1.5 ).
- Confirm integration path (REST API, Apache Arrow,
pip install jua, unified schema across 25+ models).
Jua serves major utilities across four continents, with sales cycles compressed to within weeks once the live benchmark runs. The numbers speak, and the objection shifts from “is this real?” to “how fast can we sign?”