Written by: Olivier Lam, Physical AI Team, Jua.ai AG
Key Takeaways for European Renewable Portfolios
- Capacity-normalized mean absolute error (cnMAE) provides a clean way to compare renewable forecast accuracy across portfolios of different sizes in European energy markets.
- ENTSO-E data shows that day-ahead and intraday wind and solar forecasts carry notable cnMAE, with errors spiking during winter storms, cloud-cover volatility, and dunkelflaute events.
- EPT-2 outperforms ECMWF HRES and Microsoft Aurora on 10 m wind, 100 m wind, and surface solar radiation across all lead times from 0–240 hours, with native 100 m wind output eliminating boundary-layer extrapolation error.
- Portfolio-level accuracy gains from EPT-2 translate directly into reduced imbalance costs, with a 1 GW wind portfolio achieving measurable annual savings and larger multi-GW portfolios scaling those economics significantly.
- Talk to the Jua team to benchmark EPT-2 head-to-head against your current provider on your own European region and asset mix.
ENTSO-E Day-Ahead and Intraday Error Ranges in Practice
ENTSO-E transparency data shows that day-ahead wind generation forecasts across European bidding zones can carry notable cnMAE, with higher errors during winter storm sequences and periods of rapid synoptic change. Day-ahead solar forecasts also show notable cnMAE, driven by cloud-cover volatility and low-sun-angle seasons. Intraday horizons (gate-closure to delivery, typically 1–4 hours) reduce absolute error but do not eliminate it. Residual uncertainty persists even at short lead times when mesoscale convection or frontal passages remain active.
| Variable | Horizon | cnMAE Range (Europe) | Primary Error Driver |
|---|---|---|---|
| Wind (onshore + offshore) | Day-ahead (12–36 h) | Notable under challenging conditions | Synoptic ramps, wake effects |
| Wind (onshore + offshore) | Intraday (1–4 h) | Notable under challenging conditions | Mesoscale frontal passages |
| Solar (utility + embedded) | Day-ahead (12–36 h) | Notable under challenging conditions | Cloud-cover transitions, aerosol loading |
| Solar (utility + embedded) | Intraday (1–4 h) | Notable under challenging conditions | Convective cloud development |
These observations align with Fraunhofer ISE energy-charts data for Germany, with both metrics deteriorating sharply during the dunkelflaute episodes documented in the section below. See how EPT-2 performs on your own region and variable by running a head-to-head benchmark against your current provider in under five minutes.
EPT-2 Measured Outperformance on Wind and Solar Variables
Jua is a foundation model and agent company. Jua for Energy is the first applied product, similar to the relationship between Anthropic and Claude Code. The Earth Physics Transformer (EPT) family is a general spatiotemporal transformer foundation model that learns governing physics directly from observational data. EPT-2 is the current deterministic flagship. EPT-2e is the ensemble variant.
arXiv:2507.09703 documents EPT-2 benchmark results against ECMWF HRES and Microsoft Aurora across the full 0–240 hour lead-time range. The headline findings relevant to European renewable portfolios appear below.
| Variable | Lead-Time Range | EPT-2 vs ECMWF HRES | EPT-2 vs Microsoft Aurora |
|---|---|---|---|
| 10 m wind speed | 0–240 h | Outperforms at every lead time (arXiv:2507.09703) | Outperforms across full range (arXiv:2507.09703) |
| 100 m wind speed | 0–240 h | Outperforms at every lead time (arXiv:2507.09703) | Outperforms across full range (arXiv:2507.09703) |
| Surface solar radiation (SSRD) | 0–240 h | Outperforms at every lead time (arXiv:2507.09703) | EPT-2 wins by default, because Aurora has no SSRD output (arXiv:2507.09703) |
EPT-2e, the ensemble variant, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time (arXiv:2507.09703). EPT-2 is benchmarked against more than 10,000 real ground stations via open-source StationBench, with no post-processing or station fine-tuning. EPT-2 outperforms leading AI weather models and traditional numerical baselines across all forecast horizons on RMSE, with high-resolution forecasts available over Europe.
The 100 m wind variable is particularly material for European offshore and onshore wind operators. Hub heights on modern turbines sit between 80 m and 160 m. The 10 m wind output used by most NWP-derived power forecasts requires a boundary-layer extrapolation step that introduces additional error. EPT-2 forecasts 100 m wind natively, which removes that extrapolation layer from the error budget. These accuracy gains become most consequential during the extreme weather patterns that stress forecast models hardest, particularly dunkelflaute events and rapid cloud-cover transitions.
Dunkelflaute Events and Cloud-Cover Volatility Impacts
Dunkelflaute, simultaneous low-wind and low-solar periods that typically last two to seven days, represent the worst-case scenario for renewable forecast accuracy. Research quantifying dunkelflaute events across Europe shows that these periods often cluster in the winter season, with northern and central regions more exposed. During dunkelflaute onset, day-ahead wind cnMAE can spike as models struggle to resolve the precise timing of the blocking high that suppresses both wind and solar output simultaneously.
Cloud-cover volatility introduces a separate error mechanism. Rapid transitions between overcast and broken-cloud states, driven by convective development or frontal cloud bands, produce solar forecast errors that differ structurally from the smooth diurnal cycle that most solar models are calibrated on. Intraday solar cnMAE during active convective days in Central Europe can reach significantly higher levels than during stable conditions.
Both error types are amplified by the fixed-step roll-forward architecture used by most AI weather peers. Aurora and comparable models are trained on a fixed 6-hour grid and roll forward in 6-hour increments, which compounds error at each step. EPT-2 forecasts at native any-Δt. The model is trained to predict at arbitrary time steps rather than rolling forward in fixed increments, which preserves skill at the sub-6-hour horizons where dunkelflaute onset and cloud-cover transitions matter most for intraday trading.
Portfolio-Level Imbalance Cost Implications
Timera Energy analysis identifies wind forecast error as the largest single contributor to imbalance volumes in renewables-heavy European power systems, with solar forecast error a significant secondary contributor. Summer periods with high embedded solar generation add uncertainty to demand forecasts and increase this impact. The analysis finds that imbalance exposure is inherently asymmetric. The imbalance price derives from the most expensive Balancing Mechanism action, so being on the wrong side of a forecast miss is structurally more expensive than the average spread implies. Timera stochastic modelling projects that the distribution of day-ahead versus imbalance price spreads will broaden as renewable penetration rises and dispatchable thermal capacity retires.
The market-sizing arithmetic is direct. Jua forecasts can carry significant P&L impact per gigawatt annually in European energy markets. To illustrate, a 1 GW wind portfolio that gains accuracy in cnMAE can save substantially per year under typical European hedging and imbalance penalty structures. Solar portfolios see even higher per-GW savings because imbalance prices tend to spike during peak demand periods when solar forecast errors matter most, so a 1 GW solar portfolio at the same accuracy gain can save more per year than its wind equivalent. These economics scale linearly with portfolio size, so operators running multi-GW fleets multiply the benefit. A 5 GW wind portfolio achieving improved cnMAE by replacing a standard NWP-derived forecast with EPT-2 represents substantial annual imbalance cost reduction.
The asymmetry identified by Timera means the downside of a missed dunkelflaute onset or a misforecast cloud-cover transition is not symmetric with the upside of a correct forecast. Accurate probabilistic forecasts, specifically ensemble outputs with calibrated uncertainty, are required to price that asymmetry into hedging and contracting decisions. EPT-2e outperformance of the 50-member ECMWF ENS mean on CRPS is directly relevant here. CRPS measures whether a probabilistic forecast assigns the right probability mass to the outcome that actually occurred.
Estimate your imbalance savings by running EPT-2 and EPT-2e on your own region and asset mix.
Run Your Own Benchmark in Five Minutes
The live benchmark on the Jua platform puts more than 25 models, including 10 proprietary AI models from the EPT family plus 15 third-party NWP and AI models such as ECMWF HRES, ECMWF ENS, ECMWF AIFS, NOAA GFS, DWD ICON, Microsoft Aurora, and GFS GraphCast, on a single surface. The workflow below runs in under five minutes and produces a defensible, head-to-head cnMAE comparison on the region and variable that matter to your book.
Step 1. Navigate to athena.jua.ai and open the Benchmarking surface. No pipeline setup is required.
Step 2. Select your region. The platform accepts any European bidding zone, country, or custom bounding box. For wind, select a zone where your portfolio has material exposure, such as North Sea offshore, German onshore, or the Iberian Peninsula. For solar, select the zone where cloud-cover volatility is your primary error driver.
Step 3. Select your variable. For hub-height wind, select 100 m wind speed. For solar, select surface solar radiation (SSRD). Both are native EPT-2 outputs, not post-processed extrapolations.
Step 4. Select your time window. For a day-ahead evaluation, set the hindcast window to the past 90 days. For a dunkelflaute stress test, include the most recent November–February period.
Step 5. Select your models. Add EPT-2 and your current provider, such as ECMWF HRES, a point-solution SaaS feed, or any of the 15 third-party models on the platform. The platform returns a head-to-head accuracy comparison in seconds. Athena, the Jua AI agent, can narrate the result, run a backtest over a longer historical window, or generate a custom widget for your desk. Typical queries resolve in about 90 seconds, and backtests complete in about five minutes.
Meteorologists who were sceptical of vendor accuracy claims often become internal champions once they run the benchmark themselves. The numbers speak clearly. Run benchmarks on your own region and variables on the Jua platform and see your forecasts head-to-head against more than 25 models in less than five minutes at athena.jua.ai.
Frequently Asked Questions
EPT-2 Coverage for Country-Level and Sub-Zonal European Markets
EPT-2 produces forecasts at native high resolution across Europe, covering all major bidding zones and countries. Jua for Energy currently delivers live power forecasts, including solar, wind onshore, wind offshore, total wind, total renewables, load, and residual load, for Germany, Great Britain, France, the Netherlands, and Belgium, with additional country coverage added on a weekly basis. The benchmarking surface on the Jua platform accepts any custom bounding box, so meteorologists and quant developers can evaluate EPT-2 accuracy on a specific offshore wind cluster, a solar farm cluster, or a national bidding zone without being constrained to predefined regional aggregates. ENTSO-E grid data is integrated directly into the platform, which enables capacity-weighted comparisons that match the actual generation mix of each zone.
Dunkelflaute cnMAE Spikes and EPT-2 Early Warning
During dunkelflaute onset, the transition into a simultaneous low-wind, low-solar blocking pattern, day-ahead wind cnMAE across Central Europe and the North Sea region can spike from baseline levels, with solar cnMAE rising in parallel as persistent stratiform cloud suppresses irradiance. The error spike arises from the difficulty of resolving the precise timing and spatial extent of the blocking high that drives both suppressions. EPT-2e, the ensemble variant, provides calibrated probabilistic forecasts that quantify the uncertainty around dunkelflaute onset timing. The spread of the ensemble members at the 3–7 day lead time gives a direct signal of how confident the model is about the transition date. Traders and asset operators can use the ensemble spread to size hedging positions ahead of the event rather than reacting after the blocking pattern is established. Divergence alerts on the Jua platform fire automatically when EPT-2e and other models disagree on the onset timing, which surfaces the trade window before the market reprices.
Interpreting cnMAE Improvements for a Trading Book
A reduction in day-ahead wind cnMAE for a 1 GW portfolio means that the average absolute forecast error falls. In European power markets operating under balancing-responsible-party frameworks, that reduction in average imbalance exposure translates to avoided imbalance costs and hedging penalties under typical price spreads. For solar, the same accuracy gain can lead to higher savings, reflecting the higher imbalance price sensitivity of solar-heavy systems during peak demand periods. The asymmetry of imbalance pricing, where the cost of being short the system during a price spike is structurally larger than the benefit of being long, means that the value of accuracy improvement is not linear. Reducing the tail of large errors, captured by RMSE, matters as much as reducing the average error, captured by cnMAE. CRPS, which evaluates the full probability distribution of a forecast rather than just its central estimate, is therefore the correct metric for evaluating ensemble forecasts used in hedging and contracting decisions.
How Jua for Energy Compares to Direct ECMWF or Research AI Feeds
ECMWF HRES delivers raw grib files twice a day via the MARS archive. AI weather research outputs from Microsoft Aurora or Google DeepMind GraphCast deliver raw model files without ensembles, hindcasts, or productised tooling. In both cases, the subscribing team builds the ingestion pipeline, the ensemble logic, the benchmarking harness, and the hindcast access themselves. Jua for Energy replaces that plumbing. The Jua platform exposes more than 25 models, including ECMWF HRES, ECMWF ENS, ECMWF AIFS, Aurora, and GraphCast, through a single REST API with Apache Arrow support and a Python SDK installable via pip install jua. EPT-2 and EPT-2e run on the same surface, benchmarked transparently against every other model. Athena, the Jua AI agent, turns a natural-language question into a briefing, a benchmark, a backtest, or a custom widget in about 90 seconds. The 7–9 a.m. manual prep routine, downloading grib files, processing them through brittle in-house pipelines, and waiting for the meteorologist briefing, compresses into a single workspace that refreshes on every new model run.
Peer-Reviewed Evidence for EPT-2 Accuracy
EPT-2 benchmark results appear in the technical report at arXiv:2507.09703, published July 2025. The evaluation methodology uses open-source StationBench, benchmarked against more than 10,000 real ground stations, with no post-processing or station fine-tuning applied to EPT-2 outputs. The report covers 10 m wind speed, 100 m wind speed, 2 m temperature, and surface solar radiation across the full 0–240 hour lead-time range, with head-to-head comparisons against ECMWF HRES, ECMWF ENS, and Microsoft Aurora. EPT-1.5 results appear in arXiv:2410.15076. Both reports are publicly accessible and the evaluation code is open-source, which enables independent replication by any meteorologist or quant developer who wants to verify the numbers on their own region and variable before procurement.
Conclusion: Turning Forecast Accuracy into P&L
Day-ahead wind and solar cnMAE values are not abstract statistics. At the portfolio level, they translate to substantial imbalance costs, which widen as renewable penetration rises and dispatchable thermal capacity retires. Dunkelflaute events and cloud-cover volatility push error ranges higher precisely when imbalance prices reach their peaks. Transparent, head-to-head benchmarking now sits within reach. The numbers are available, the methodology is open-source, and the evaluation runs in five minutes on the Jua platform.
The performance advantage documented earlier holds across every variable that drives a European renewable P&L, at every forecast horizon that matters for day-ahead and intraday trading. EPT-2e beats the 50-member ECMWF ENS mean on RMSE and CRPS at virtually every lead time. These results are verifiable on your own region and variable, against your own current provider, in less than five minutes on the Jua platform.
Schedule a portfolio-specific walkthrough and run the benchmark on your book before the next dunkelflaute onset.