A complete guide to Steadysun’s renewable energy forecasting methodology, from satellite nowcasting to multi-model NWP, AI calibration, probabilistic outputs, and the full parameters catalogue.
Solar power forecasting is the process of predicting the electricity output of a photovoltaic over a defined time horizon, from the next minute to 15 days ahead.
Wind energy forecasting applies the same core discipline to turbines, replacing irradiance physics with wind speed and turbulence dynamics.
Both disciplines share a common problem: weather is uncertain, and that uncertainty costs money. The goal of any serious renewable energy forecasting system is to shrink that uncertainty as much as physically possible, then quantify what remains so operators can act on it.
Since 2013, Steadysun has been refining how it forecasts solar and wind power, fusing live data from 5 geostationary satellites, 20+ numerical weather prediction (NWP) models, and real-time plant measurements. This gives deterministic and probabilistic forecasts from 0 to 15 days ahead, updated as frequently as every 5 minutes, delivered worldwide via API or SFTP.
There are three broad families of methods behind this fusion. In practice, the most accurate systems use all three together.
Numerical Weather Prediction (NWP) models solve the equations of atmospheric dynamics on a 3D grid, starting from observed initial conditions and stepping forward in time.
Models like ECMWF’s IFS, NOAA’s GFS, Météo-France’s ARPEGE, or the regional WRF (which can reach 1 km resolution) are the backbone of any serious day-ahead or medium-range forecast.
For solar, NWP outputs (cloud fraction, aerosol optical depth, humidity) feed into irradiance models.
For wind, NWP wind fields are downscaled to hub height using boundary-layer physics or statistical transfer functions.
The limitation: NWP models have systematic biases that vary by location, season, and weather regime. No single model is best everywhere, all the time.
D+1–D+15 (day-ahead & medium-range)
up to every hour
10–50 km (global)
5–10 km (regional high res)
1–5 km (regional high res)
Down to 1 km (regional WRF)
For the next few minutes to a few hours, live satellite imagery outperforms NWP models, which simply can’t update fast enough to capture a cloud passing over a plant right now.
The principle: track cloud motion across multiple spectral channels (visible, near-infrared, thermal infrared) between successive images, then extrapolate cloud cover and irradiance forward in time.
The limitation: this approach loses accuracy fast beyond a few hours, since cloud motion extrapolation can't anticipate new cloud formation or dissipation, that's exactly where NWP models take over.
0–30 min (nowcasting) and 30 min–3 h (intraday)
every 5–15 minutes
500 m–3 km
Statistical approaches, from simple persistence (assuming the next hour looks like the current hour) to ARIMA time series models, work well at very short horizons where the atmosphere hasn’t changed much.
Machine learning goes further. Neural networks, gradient boosting, and ensemble methods can learn non-linear relationships between NWP outputs and observed power, correcting biases that physics-based models miss. They need training data (typically at least one year of co-located production and weather records) but once trained, they adapt continuously as new data arrives.
The limitation: purely data-driven models can struggle in rare or extreme weather situations that weren't well-represented in the training set.
Steadysun’s approach is neither purely physical nor purely statistical. It’s a hybrid multi-model engine that combines the strengths of both.
The core idea: run 20+ leading NWP models in parallel, then use AI to select, weight, and merge their outputs into a single, more reliable forecast. The AI assigns dynamic weights to each NWP model based on its recent performance over a rolling window.
This is conceptually similar to what the forecasting community calls a “meta-forecast” or ensemble post-processing, the same principle behind ECMWF’s ensemble prediction system, but applied operationally at the plant level and updated every 5 minutes.
The result is a forecast that’s more robust than any individual NWP model, because the errors of different models tend to be uncorrelated. When one model is wrong, others compensate.
Key facts up front :
| Forecast horizons | 0–30 min (nowcasting) and 30 min–3 h (intraday) for solar only D+1–D+15 (day-ahead & medium-range) for both solar and wind |
|---|---|
| Frecuencia de actualización | Every 5–15 min (satellite-based) / up to every hour (NWP-based) |
| NWP models aggregated | 20+ including IFS, AIFS, GFS, CAMS, ICON, ICON-EU, ICON-D2, ICON-2I, ARPEGE, ARPEGE-EU, AROME, HRRR, HRDPS, GDPS, WRF, HARMONIE-NEA and more. |
| NWP spatial resolution | 10–50 km (global) 5–10 km (regional high resolution) 1–5 km (regional high resolution) down to 1 km (regional WRF) |
| Geostationary satellites | 5 systems including GOES West, GOES East, Meteosat 0°, Meteosat IODC, y Himawari. |
| Satellite spatial resolution | 500 m – 3 km |
| Satellite update cadence | Every 5–15 minutes |
| Technologies covered | PV (fixed, tracking, bifacial), CSP, onshore wind, offshore wind |
| Probabilistic output | 11 quantiles: P00, P10, P20 … P90, P100 |
As seen above, no single technology covers the full forecasting timeline. At Steadysun, we split it across two engines, each taking the lead at a different horizon: Steadysat for the short range, Steadymet for the long range.
t-10min – 30 minutes
Nowcasting
(Steadysat)
30 minutes – 6 hours
Intraday
(Steadysat + Steadymet blend)
6 hours – D+15
Day-ahead & Medium range
(Steadymet)
Steadysat is Steadysun’s implementation of the satellite-based approach described above.
At very short time scales, the dominant uncertainty is cloud motion, and NWP models are too coarse and too slow to capture a cloud front moving across a solar farm in the next 15 minutes. Steadysat solves this using cloud motion vectors (CMV) derived from geostationary satellite imagery.
By tracking cloud displacement between successive satellite images (refreshed every 5 to 15 minutes at 500 m–2 km spatial resolution) Steadysat extrapolates where clouds will be in the next 30 minutes with high spatial precision.
The 5 satellites feeding Steadysat are:
Steadysat also ingests real-time in-situ measurements (irradiance sensors, pyranometers, power meters, SCADA data) to continuously calibrate its outputs at the plant level.
Beyond 30 minutes, cloud motion vectors alone lose predictive skill as cloud systems evolve and dissipate. Steadysat extends its horizon to 6 hours by blending CMV-based extrapolation with the forecast from Steadymet, our proprietary hybrid multi-model engine, weighting each source according to its demonstrated skill at each time step.
This blended approach is updated every 5 to 15 minutes, making it ideal for intraday market participation, storage dispatch, and island grid balancing.
For horizons beyond 6 hours, Steadymet takes over. It aggregates output from more than 20 global and regional NWP models and uses machine learning to eliminate the systematic biases of each individual model (the same principle described above in the hybrid multi-model approach), here run continuously in production for every plant.
Global NWP models such as GFS and ECMWF run at 10–25 km horizontal resolution. For sites where local topography, coastal effects, or island microclimates matter, Steadysun can deploy the regional WRF (Weather Research and Forecasting) model at 1 km resolution, implemented and optimised by our in-house team of meteorologists.
Steadymet delivers forecasts up to 15 days ahead, updated up to every hour.
Before any forecast can be tailored to your site, you provide the technical details needed to build a high-fidelity digital representation of your asset.
Once your site is set up, forecasting can begin.
Every forecast cycle start with data ingestion from multiple independent sources:
Satellite observations (Steadysat)
NWP model output (Steadymet)
Ground measurements
Once your site is onboarded, where historical production data is available (typically at least several months of SCADA or meter data), Steadymet and Steadysat are trained on the relationship between observed weather and observed power. The system learns the plant’s specific behaviour and corrects systematic biases (soiling losses, inverter clipping, wake effects, local aerosol climatology).
For new installations without production history, Steadysun’s Hindcast service can provide a solution. Hindcast runs the current Steadymet and Steadysat models on historical weather periods at your specific site, simulating what the forecasting system would have predicted in the past. This establishes a rigorous performance baseline (including MAE, RMSE, and full quantile distributions) before live operation begins. It’s particularly valuable for financial model validation, PPA negotiations, and BESS sizing.
During live operation, Steadymet and Steadysat continuously ingest real-time plant measurements to update their site-specific correction layers. If the forecast is running 5% high this morning, the system adjusts. Machine learning algorithms detect emerging biases, for example, a systematic underestimation of morning irradiance due to coastal fog, and correct them dynamically.
This continuous learning loop means forecast accuracy typically improves over the first weeks and months of operation as the models accumulate site-specific knowledge.
Raw meteorological forecasts are converted into AC or DC power output using the plant’s physical model, with high-resolution topographical corrections down to 90 m resolution.
Forecasts are then delivered via API, SFTP, CSV, NetCDF or Dashboard, along with their uncertainty metrics, at time steps down to 1 minute.
Quality is tracked continuously using three standard metrics:
Put our forecasts to the test! Easily integrate our data into your tools and see the impact firsthand. Optimize two plants free with Expert+ features – no commitment !
Weather is inherently uncertain. Instead of relying on a single forecast, Steadysun captures that uncertainty with probabilistic forecasts that show the full range of possible outcomes.
Each forecast point includes 11 quantiles (P00, P10, P20, P30, P40, P50, P60, P70, P80, P90, P100), giving you a clearer picture of what to expect and how uncertain the forecast is.
What these mean in practice:
In practice, the P10–P90 interval gives a robust 80% confidence band. The P00 and P100 bounds represent the plausible worst and best cases, useful for stress-testing storage dispatch strategies or estimating maximum imbalance exposure.
These quantiles are generated using a combination of three complementary sources.
Propagating uncertainty from the spread across 20+ NWP models.
Learning the empirical distribution of forecast errors at each site and horizon.
Using cloud cover variability within the satellite footprint to modulate short-term uncertainty.
While quantiles provide a range of possible outcomes for each hour, they don’t show how these outcomes might evolve over time. Scenario generation creates full, realistic time-series simulations, complete and internally consistent trajectories that span the entire forecast horizon. This means you don’t just see isolated worst- or best-case moments; you see how conditions could unfold hour by hour.
In 2024, a developer planning a hybrid PV+BESS plant in northern Mauritius used Steadysun’s probabilistic hindcast to size their battery and estimate imbalance penalties before commissioning.
The result was a 15–31% improvement over persistence forecasts across different time horizons, and full compliance with the local Grid Code.
Ignoring uncertainty doesn’t make it go away. It just means someone else is bearing it (usually the grid).
The table below lists all weather and production parameters available.
| Parameter | Nombre corto | Descripción | Unidad |
|---|---|---|---|
| AC power | pac | Potencia AC solar en la salida del inversor | W |
| DC power | pdc | Solar DC power at array output | W |
| Global Horizontal Irradiance | ghi | Total solar irradiance on a horizontal surface | W·m⁻² |
| Direct Normal Irradiance | dni | Direct beam irradiance on a surface normal to the sun | W·m⁻² |
| Diffuse Horizontal Irradiance | dhi | Diffuse irradiance on a horizontal surface | W·m⁻² |
| Global Tilted Irradiance | gti | Total irradiance on a tilted (panel) surface | W·m⁻² |
| Beam Normal Irradiance | bni | Beam irradiance component, normal incidence | W·m⁻² |
| Beam Horizontal Irradiance | bhi | Beam irradiance on a horizontal surface | W·m⁻² |
| Photosynthetically Active Radiation | par | Solar radiation in the 400–700 nm band | W·m⁻² |
| Parameter | Nombre corto | Descripción | Unidad |
|---|---|---|---|
| High cloud cover | hcc | Fraction of sky covered by high-level clouds | [0–1] |
| Medium cloud cover | mcc | Fraction of sky covered by mid-level clouds | [0–1] |
| Low cloud cover | lcc | Fraction of sky covered by low-level clouds | [0–1] |
| Total cloud cover | tcc | Total fraction of sky covered by clouds | [0–1] |
| Precipitation rate | mtpr | Liquid water precipitation rate | mm·h⁻¹ |
| Relative humidity | rh2m | Relative humidity at 2 m above ground | % |
| Specific humidity | q | Mass of water vapour per unit mass of moist air | g·kg⁻¹ |
| Total column water vapour | tcwv | Integrated atmospheric water vapour column | kg·m⁻² |
| Parameter | Nombre corto | Descripción | Unidad |
|---|---|---|---|
| Ambient temperature | t2m | Air temperature at 2 m above ground | °C |
| Dewpoint temperature | td | Dewpoint temperature at 2 m | °C |
| PV cell temperature | t_cell | Estimated PV module cell temperature | °C |
| Parameter | Nombre corto | Descripción | Unidad |
|---|---|---|---|
| Wind speed at 10 m | 10m_wind_speed | Mean wind speed 10 m above ground | m·s⁻¹ |
| Wind direction at 10 m | 10m_wind_direction | Wind direction at 10 m (0° = North) | ° |
| Wind gust at 10 m | 10m_gust | Maximum wind gust speed at 10 m | m·s⁻¹ |
| Wind speed at hub height | hhws | Wind speed at turbine hub height (25–100 m) | m·s⁻¹ |
| Mean sea level pressure | mlsp | Atmospheric pressure reduced to sea level | hPa |
| Surface pressure | msp | Atmospheric pressure at the surface | hPa |
If your workflow requires additional variables, our team can add them to your forecast outputs to meet your specific operational needs.
Ofrecemos varias opciones de integración adaptadas a sus sistemas:
Our solution is technology-agnostic and covers:
Use cases span the full project lifecycle:
| Phase | Use case |
|---|---|
| Investigación y innovación | Forecasting method benchmarking, academic collaboration, access to historical datasets |
| Development | Hindcast-based performance baseline, PPA validation, BESS sizing |
| Construction | Pre-commissioning model setup, satellite calibration |
| Operations | Real-time dispatch, grid code compliance, O&M scheduling |
| Trading | Day-ahead bidding, intraday optimisation, imbalance cost reduction |
| Gestión de redes | Reserve scheduling, curtailment management, frequency regulation |
| Self-consumption | Load-matching optimisation, battery charge/discharge scheduling, export minimisation |
Respuestas a tus preguntas
The core methodology is the same: NWP models + AI multi-model engine + probabilistic output. The physics are different.
Solar power forecasting is driven by irradiance and cloud cover; wind energy forecasting is driven by wind speed and direction at hub height, atmospheric stability, and turbulence. Wind is also more sensitive to local terrain effects; a ridge or valley can accelerate or deflect wind in ways that a global NWP model at 10–25 km resolution can’t fully resolve, which is why site-specific calibration is especially important for wind. Solar has an additional intraday tool (Steadysat, using geostationary satellites) that has no direct equivalent for wind.
Steadysat is Steadysun’s satellite-based nowcasting and intraday forecasting product. It processes live imagery from 5 geostationary satellites to track cloud motion and forecast solar irradiance up to 6 hours ahead, updated every 5–15 minutes.
Steadymet is the multi-model NWP-based product for day-ahead and medium-range forecasting (up to 15 days), aggregating 20+ weather models and applying AI-based bias correction.
The two products are complementary and together cover the full 0–15 day horizon.
Forecast accuracy is measured using MAE, MBE, and RMSE, and depends on three factors:
Short-horizon forecasts (intraday, up to 3 h) are generally more accurate than day-ahead or medium-range forecasts.
What we can say: in one deployment, better solar forecasting cut grid code penalties by 50% and curtailment losses by 75% — results specific to that site and not a guarantee for every project.
Yes. Steadysat’s global satellite coverage and Steadymet’s multi-model approach both operate worldwide. Steadysun has specific experience with island and insular grids, including Tahiti, Mauritius, and New Caledonia, where forecasting challenges are amplified by limited grid inertia and high renewable penetration.
The regional WRF model at 1 km resolution is particularly effective in these environments.
Every Steadysun forecast includes 11 probabilistic quantiles (P00 to P100), generated from the spread across NWP ensemble members and from site-specific error distributions.
These quantiles allow users to set risk-adjusted bidding strategies, size storage systems under different confidence levels, and comply with grid codes that require uncertainty-aware production schedules.
Contáctanos, we will be happy to answer you!
English (UK) English (UK)
Français
EspañolRecibe cada trimestre análisis del sector y novedades de nuestros productos
