Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: July 12, 2026
Key Takeaways for Energy Desks
- The ECMWF AIFS v2 upgrade from May 2026 adds wave and snow cover forecasts and improves medium-range skill for 100 m wind and surface solar radiation, which directly affect renewable energy P&L.
- AIFS Single and AIFS ENS now run alongside the physics-based IFS, yet still depend on external platforms for live benchmarking, natural-language analysis, and power-forecast integration.
- EPT-2 outperforms ECMWF HRES on every lead time for 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation, and EPT-2e beats the 50-member ECMWF ENS mean on RMSE and CRPS at virtually every lead time.
- Energy traders gain the most value when they run multiple models under one schema with frequent updates, 15-minute power forecasts, and roughly 90-second Athena briefings, which the Jua platform delivers natively.
- Book a demo with Jua to benchmark AIFS live against EPT-2 and 23 other models on your region and variables.
How the ECMWF AIFS Weather Model Works
AIFS is ECMWF’s data-driven forecasting system and sits apart from the institution’s traditional numerical weather prediction approach. Numerical weather prediction decomposes the atmosphere into three-dimensional grid cells and solves differential equations inside each one. AIFS replaces that physics solver with a learned model trained on historical atmospheric states.
The architecture follows an encoder-processor-decoder design. The encoder and decoder use attention-based graph neural networks, and the processor is a transformer that applies sliding window attention. Input and output operate on the N320 reduced Gaussian grid at approximately 0.25° horizontal resolution, while the internal processor uses a coarser ~1° grid with 16 layers.
Training data spans two phases. The model first pre-trains on the Copernicus ERA5 reanalysis from 1979 to 2022. It then undergoes rollout fine-tuning on ECMWF operational IFS analysis from 2016 to 2022. Lead time extends to 10 days for AIFS Single and 15 days for AIFS ENS. Forecasts arrive as grib files via ECMWF’s standard dissemination channels.
Inference runs quickly. A 10-day AIFS Single forecast completes in approximately 2 minutes 30 seconds on a single NVIDIA A100 40 GB GPU, which allows dissemination ahead of the physics-based model chain. This speed advantage becomes operationally meaningful with the May 2026 v2 release, which broadens AIFS coverage and improves robustness for energy-relevant variables.
What Changed in the May 2026 AIFS v2 Release
AIFS v2 went live on 12 May 2026, upgrading both the Single and Ensemble configurations simultaneously. Key changes in v2 include:
- ECMWF’s first data-driven wave forecasts with 11 new wave-related variables, which show substantial improvement in medium-range wave skill versus IFS Cycle 50r1.
- ECMWF’s first data-driven snow cover forecasts, where predicted snow cover fraction tracks observations more closely than IFS Cycle 50r1.
- Fine-tuning on a dataset that includes both operational analysis and prototype IFS Cycle 50r1 data, which reduces sensitivity to IFS cycle upgrades, a known vulnerability of earlier AIFS versions and external AI models.
The prior operational version, AIFS Single 1.1.0, went live on 27 August 2025 to correct a precipitation forecast issue. That release also introduced 100 m wind components and surface solar radiation downwards as prognostic variables, which both matter directly for renewable energy forecasting.
On the ensemble side, the operational 51-member AIFS ENS (AIFS-CRPS) has run daily at 00/06/12/18 UTC since 1 July 2025. It produces forecasts to 15 days at approximately 30 km horizontal resolution and therefore updates 4 times per day. CRPS, or Continuous Ranked Probability Score, measures the accuracy of probabilistic forecasts by comparing the full forecast distribution against observed outcomes, where lower CRPS indicates better probabilistic skill.
Energy teams can run live benchmarks on their own region and variables at athena.jua.ai.
Accuracy Benchmarks Against NWP and Other AI Models
RMSE, or Root Mean Square Error, measures the average magnitude of forecast errors, so lower values indicate higher deterministic accuracy. CRPS evaluates probabilistic forecast quality across the full ensemble distribution.
AIFS outperforms the operational IFS by 12 to 24 hours in anomaly correlation skill during the medium range from 3 to 10 days. Overall skill improvements of 4 to 6 percent across all variables, lead times, and pressure levels appear relative to the previous AIFS version, with up to 12 percent improvement in precipitation skill.
Against other AI models, the picture looks more competitive. Raw ECMWF IFS ensemble forecasts and raw AIFS ensemble forecasts have been compared for 10 m wind speed across all lead times from 24 hours to 360 hours in a global study of 9,246 SYNOP stations from July to November 2025. Post-processing with ensemble model output statistics narrows this gap, and differences between post-processed IFS and AIFS remain statistically significant only through day 4.
EPT-2, Jua’s general physics foundation model fine-tuned for atmospheric prediction, maintains a performance advantage over ECMWF HRES across the full 0 to 240 hour range, as documented in arXiv:2507.09703. The ensemble variant, EPT-2e, extends the same pattern to probabilistic forecasts, as documented in arXiv:2410.15076.
Single vs Ensemble Performance for Traders
AIFS operates in two distinct configurations. AIFS Single is trained with a mean-squared error loss function and produces a single deterministic forecast trajectory. AIFS ENS is trained probabilistically and runs 51 members, initialized 4 times per day, with a 15-day horizon at approximately 30 km resolution.
For energy trading, the ensemble configuration carries more weight than the deterministic one. Probabilistic forecasts allow traders to quantify uncertainty, size positions around forecast spread, and see when models converge or diverge on a key variable. A single deterministic trajectory provides no uncertainty estimate and therefore no basis for probabilistic position sizing.
EPT-2e, Jua’s ensemble variant documented in arXiv:2507.09703, beats the 50-member ECMWF ENS mean on RMSE and CRPS at virtually every lead time and extends to a 60-day ensemble horizon. It also updates 4 times per day, which pushes beyond AIFS ENS’s 15-day ceiling while matching its update cadence.
Teams can run live benchmarks on their own region and variables at athena.jua.ai.
AIFS Compared with GraphCast and Aurora
ECMWF discontinued real-time operation of external ML models, including Pangu-Weather, GraphCast, Aurora, and FourCastNet, with the implementation of IFS Cycle 50r1 because of performance degradation caused by changes in initial conditions. GraphCast and Aurora showed reduced performance under the new analysis cycle, and Aurora and PanguWeather also lack precipitation forecasts, which limits their operational utility. AIFS v2 was fine-tuned specifically to mitigate this sensitivity.
GraphCast and FourCastNet have been superseded by probabilistic models from their respective research groups, which reflects a broader shift toward ensemble approaches for operational suitability. Neither GraphCast nor Aurora ships a productised ensemble equivalent that fits operational energy trading workflows.
EPT-2 beats Microsoft Aurora on 10 m wind, 100 m wind, and 2 m temperature across the full 0 to 240 hour range. Aurora provides no surface solar radiation downwards output, which creates a critical gap for solar generation forecasting. EPT-2 inference runs approximately 25 percent faster than Aurora, was trained on 8 H100 GPUs over 10 days versus Aurora’s 32 A100 GPUs over 18 days, and costs roughly 0.20 to 15 dollars per simulation on a single GPU compared with 1,000 to 20,000 euros for a traditional NWP run.
How Well AIFS Fits Energy Trading Workflows
The economic case for forecast accuracy improvement is concrete for energy desks. A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves approximately 1.5 million euros per year under typical hedging and penalty structures. A 1 GW solar portfolio at the same accuracy gain saves approximately 3 million euros per year, and multi-GW portfolios scale these figures linearly.
AIFS v2 adds 100 m wind and surface solar radiation, which are the two variables most directly tied to wind and solar generation forecasting. This change makes AIFS more relevant to energy trading workflows than earlier versions. The remaining gaps sit on the operational side. AIFS updates only 4 times per day, which limits responsiveness to rapidly evolving weather events. Its grib file output requires in-house ingestion pipelines that most trading desks must build and maintain. Beyond the raw forecast data, AIFS provides no natural-language analysis layer, no live cross-model benchmarking, and no power forecast surface, so teams must construct these capabilities separately.
Jua for Energy addresses these gaps directly. EPT-2 RR updates up to 24 times per day, which gives traders a much denser intraday signal. Actual-generation power forecasts refresh every 15 minutes with a 48-hour horizon. Athena converts a natural-language question into a briefing, benchmark, backtest, or custom widget in approximately 90 seconds. The Python SDK installs via pip install jua and exposes more than 25 models, including AIFS, through a single schema with Apache Arrow support for large payloads. Jua’s models can natively forecast at up to 5 km resolution, and the Jua platform product can deliver output at up to 1 km resolution.
Book a demo to see how Jua for Energy integrates AIFS alongside EPT-2 in a single operational workspace.
Head-to-Head Model Comparison for Energy Use
| Model | Deterministic Accuracy vs HRES | Ensemble Capability | Update Frequency | Energy-Trading Workflow Features |
|---|---|---|---|---|
| ECMWF AIFS (v2, May 2026) | 12 to 24 hour anomaly correlation skill gain versus IFS in the medium range from 3 to 10 days and 4 to 6 percent overall skill improvement versus the prior AIFS version | 51 members, 15-day horizon, approximately 30 km resolution, with raw AIFS ENS and raw IFS ENS compared for 10 m wind CRPS across lead times from 24 hours to 360 hours | 4 times per day at 00, 06, 12, and 18 UTC | 100 m wind and surface solar radiation variables added in v1.1.0, grib dissemination, and no native benchmarking, briefing, or agent layer |
| Microsoft Aurora | Loses to EPT-2 on 10 m wind, 100 m wind, and 2 m temperature across 0 to 240 hours and provides no surface solar radiation output | No productised ensemble and ECMWF discontinued real-time Aurora operation because of performance degradation under IFS Cycle 50r1 | Typically 4 times per day research cadence with no productised operational schedule | No native power forecast, briefing, benchmarking, or agent layer and only raw model output |
| Google DeepMind GraphCast | Reduced performance under IFS Cycle 50r1 and superseded by probabilistic successor models | No productised ensemble equivalent for operational use | Typically 4 times per day research cadence with no productised operational schedule | No native power forecast, briefing, benchmarking, or agent layer and only raw model output |
| EPT-2 / EPT-2e (Jua) | Outperforms ECMWF HRES on every lead time for 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation across 0 to 240 hours | EPT-2e beats the 50-member ECMWF ENS mean on RMSE and CRPS at virtually every lead time, offers a 60-day ensemble horizon, and updates 4 times per day | EPT-2 updates 4 times per day, EPT-2 RR updates up to 24 times per day, and actual-generation power forecasts update every 15 minutes | More than 25 models on one platform, Athena natural-language briefings in roughly 90 seconds, live benchmarking, Python SDK, Apache Arrow support, native 5 km resolution, 1 km product resolution, and power forecasts in 5 countries |
Frequently Asked Questions
What is the most accurate AI weather model for energy trading?
Accuracy depends on the variable and lead time, but a clear pattern emerges for energy P&L variables. For 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation, EPT-2 outperforms ECMWF HRES on every lead time across the full 0 to 240 hour range, as documented in peer-reviewed technical reports on arXiv. EPT-2e, the ensemble variant, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time. AIFS v2 shows meaningful improvement over IFS in the medium range and now includes 100 m wind and surface solar radiation variables, which makes it more relevant to energy applications than earlier versions. The most defensible operational approach for energy trading is to run multiple models simultaneously and benchmark them continuously against ground-truth observations, which the Jua platform’s live benchmarking surface enables across more than 25 models.
How does AIFS ensemble skill compare with EPT-2e?
AIFS ENS runs 51 members at approximately 30 km resolution with a 15-day horizon and updates 4 times per day. Raw AIFS ENS underperforms raw IFS ENS on 10 m wind CRPS across all lead times from 24 hours to 360 hours in global verification against SYNOP stations. Post-processing with ensemble model output statistics narrows this gap, and statistically significant differences remain only through day 4. EPT-2e beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time, extends to a 60-day ensemble horizon, and updates 4 times per day. For energy trading applications that require probabilistic skill at medium and extended ranges, such as wind ramp detection, solar generation uncertainty quantification, and gas demand spread, EPT-2e provides a deeper ensemble signal than AIFS ENS at comparable update frequency.
Can energy teams replace their ECMWF subscription with AIFS?
Energy teams should treat AIFS as part of their ECMWF subscription rather than a replacement for it. AIFS is an ECMWF product that is disseminated through ECMWF’s standard channels and sits inside the ECMWF subscription. The more relevant question concerns the plumbing around that subscription, including the in-house grib ingestion pipeline, the manual benchmarking, the morning briefing analyst, and the dashboard stitching. Jua for Energy replaces that surrounding infrastructure. Serious customers keep their ECMWF subscription and run Jua for Energy alongside it. AIFS itself runs natively on the Jua platform in the same workspace as EPT-2, EPT-2e, ECMWF HRES, ECMWF ENS, and more than 20 other models under a single schema and a single API.
Where can meteorologists run live benchmarks on AIFS versus EPT-2?
The Jua platform’s live benchmarking surface at athena.jua.ai puts more than 25 models, including AIFS, EPT-2, EPT-2e, ECMWF HRES, ECMWF ENS, Aurora, GraphCast, NOAA GFS, DWD ICON, and others, on a single screen. A meteorologist selects any region, any variable, and any time window, and the platform returns a head-to-head accuracy comparison in seconds, verified against more than 10,000 real ground stations via Jua’s open-source StationBench methodology with no post-processing or station fine-tuning. Backtests against years of historical forecasts run in approximately 5 minutes via Athena. The same surface remains available post-procurement for ongoing model surveillance as AIFS and EPT model versions update.
Conclusion for Energy Trading Teams
The May 2026 AIFS v2 upgrade represents a meaningful step for operational forecasting. Improved medium-range skill, the addition of wave and snow cover forecasts, and reduced sensitivity to IFS cycle changes make AIFS a more complete operational model than its predecessors. The inclusion of 100 m wind and surface solar radiation in v1.1.0, retained and extended in v2, directly addresses the two variables that matter most for renewable energy trading.
AIFS v2 does not, however, resolve the workflow problem for trading desks. Energy trading teams still need a single workspace that runs AIFS alongside EPT-2, EPT-2e, and 23 other models, with up to 24 daily updates via EPT-2 RR, 15-minute actual-generation power forecasts, Athena-driven natural-language briefings, and a Python SDK that stands up in days rather than a quarter.
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. Serious customers keep their ECMWF subscription and run Jua for Energy alongside it.
Book a demo to run live benchmarks on AIFS, EPT-2, and 23 other models on your own region and variables and see the numbers before the market does.