Open-Source Tool Boosts Intraday Solar Forecasts

What the framework does – instant answer

The new open‑source framework can predict national‑scale PV output for the next two hours with a typical error below 10% on more than 80% of days, giving grid operators a reliable short‑term view of solar generation.

Developed by researchers from TU Delft and Bern University of Applied Sciences, the system combines satellite‑derived solar irradiance, optical‑flow motion tracking and physics‑based numerical weather prediction (NWP). All code and data are publicly available, so utilities, aggregators and even hobbyists can run the forecasts on their own servers.

---

How it works – the three‑layer engine

1. Satellite‑based deep learning – The team uses the HANNA dataset, which provides 15‑minute surface solar irradiance (SSI) images over Switzerland. These images are turned into clear‑sky index (CSI) fields, normalising for weather‑free conditions.
2. Optical‑flow motion estimation – By feeding four consecutive CSI images (the past hour) into an optical‑flow model, the system learns how cloud patterns move and evolve.
3. Numerical weather prediction (NWP) bias correction – Outputs from a physics‑based NWP model (IFS‑ENS) are corrected with the satellite‑driven predictions, producing a hybrid forecast that leverages both data‑driven and physical insights.

Each of the six tested forecasting pipelines – SolarSTEPS, SolarSTEPS‑pa, IrradianceNet, SHADECast, IFS‑ENS and a bias‑corrected IFS‑ENS – receives the same CSI inputs and generates eight future CSI frames (covering the next two hours). Those frames are converted back to SSI and fed into an XGBoost model trained on the historical production of 6,434 Swiss PV installations, turning irradiance forecasts into power forecasts.

---

Performance results – the numbers speak

• Accuracy – The deterministic IrradianceNet model achieved the lowest root‑mean‑square error (RMSE), while the probabilistic SolarSTEPS and SHADECast ensembles offered the best‑calibrated uncertainty bands.
• Satellite advantage – At short lead times (up to 30 minutes) satellite‑based approaches outperformed the pure NWP IFS‑ENS model, especially in complex terrain.
• National‑scale robustness – Across 2019‑2020, daily total PV generation was forecast with relative errors under 10% on 82 % of days, demonstrating that the method works reliably for an entire country.

The authors stress that forecast skill declines with elevation, a known challenge for mountainous regions, but the overall performance makes the framework ready for operational use.

---

Why it matters for grid operators – the bottom line

Knowing the exact amount of solar power that will be injected into the grid a few hours ahead lets operators schedule backup generators, storage dispatch and demand‑response actions more efficiently. As Angela Meyer (TU Delft) explained, “you can act strategically when there are surpluses or shortages of energy and reduce electricity costs.”

When forecasts are accurate, utilities avoid costly ad‑hoc measures such as firing up gas turbines at short notice, which can shave several percent off overall system operating costs.

---

What it means for Israel – a quick calculation

Israel’s residential solar tariff is about ₪0.48 /kWh and a typical 10 kWp rooftop system yields roughly 17,000 kWh / year in the central region, worth ≈ ₪8,160 per year (see verified Israeli facts). If an intraday forecast reduces the need for expensive peaker‑plant dispatch by just 5 %, the saved electricity cost would be about ₪408 per year per household. Over a 25‑year system life, that adds up to ≈ ₪10,200 in avoided costs – a tangible boost to the already attractive payback of ≈ 3.9 years for a typical installation.

For large‑scale solar farms, the same percentage improvement translates into megawatts of avoided ancillary services, helping Israel meet its 30 % renewable target by 2030 while keeping consumer bills low.

---

Future outlook – open data, open impact

Because the framework is released under an open‑access licence, any stakeholder can improve the models, add local satellite feeds or adapt the code to Israel’s climate. The authors intend to continue developing the system, exploring longer forecast horizons and higher‑resolution satellite data, which may further improve accuracy.

As solar capacity expands in Israel, tools that turn clouds into reliable numbers will be key to a smoother, cheaper, and greener grid.