AI Wind and Solar Forecasting in 2026: GraphCast, Pangu-Weather and the Neural NWP Era

The single biggest technical event in energy forecasting between 2023 and 2026 has been the arrival of neural numerical weather prediction at production quality. GraphCast from Google DeepMind, Pangu-Weather from Huawei, FourCastNet from NVIDIA and the newer GenCast ensemble model have moved from research curiosities in 2023 to operational inputs to renewable forecasting pipelines in 2026, beating the physics-based baselines that ECMWF and NOAA have refined over decades — and running in a fraction of the compute. This post walks through what that means for the wind and solar forecasting vendors and the operators who depend on them.

The neural NWP revolution#

GraphCast, published in late 2023 and operationalised through 2024-2025, demonstrated that a graph neural network trained on forty years of reanalysis data could match or beat the ECMWF IFS deterministic forecast at ten-day horizons while running on a single TPU in roughly a minute. Pangu-Weather from Huawei produced comparable results around the same time. GenCast, the ensemble successor to GraphCast, addressed the calibration concerns about deterministic neural models and is now the model that several commercial forecasters quietly run alongside the IFS ENS ensemble.

By 2026, ECMWF, the UK Met Office and NOAA have all integrated some version of these models into their operational suites — not always as the lead forecast but as parallel runs whose performance is now reliably tracked. The commercial vendors who serve wind and solar operators have been more aggressive, often running blended ensembles where neural NWP outputs sit alongside the traditional GFS and IFS feeds.

The vendor map#

DNV’s WindGEMINI and the broader DNV renewables analytics suite remain the standard for asset-performance and longer-horizon forecasting in the wind industry. DNV’s depth on turbine physics — twenty years of fleet-performance data, blade-pitch dynamics, wake modelling — is what differentiates its forecasts from generic ML approaches. IBM’s HyperWatch (built on the older Deep Thunder and Watson lineage) is a meaningful player at utility-scale customers, particularly in North America and the Middle East.

Vaisala, originally a weather instrumentation company, has built a strong global renewable forecasting business by combining its hardware-derived measurement archive with modern ML. Solcast, now part of DNV after the 2022 acquisition, dominates solar irradiance forecasting with a satellite-derived global product that virtually every large solar operator subscribes to. Specialist competitors include Meteomatics, Reuniwatt, Clir Renewables, Utopus Insights (Vestas-owned), and Power Factors. Most large operators run two or three forecasts in blend.

Where the accuracy gains actually came from#

The headline accuracy gain from neural NWP at the global atmospheric level translates unevenly into renewable forecasting accuracy at the asset level. Wind forecasts depend on hub-height wind speed, which is interpolated from the model’s lowest atmospheric levels — and the interpolation, terrain correction and wake adjustment matter as much as the underlying global forecast. Solar forecasts depend on irradiance at panel level after accounting for cloud cover, aerosols and ground reflectance. The vendors who got the largest improvements in 2024-2026 are the ones who replaced their statistical post-processing with ML models trained on plant-specific historical performance — not the ones who simply swapped GFS for GraphCast at the top of the pipeline.

At the hourly horizon, the operational metric most ISOs care about, the better commercial wind forecasts in 2026 routinely deliver mean absolute errors below seven percent of rated capacity at large fleets, versus ten to fifteen percent five years ago. Solar errors have shrunk less but ramp-event prediction — the moment a cloud front crosses a plant — has improved substantially.

The intraday and ramp problem#

Day-ahead forecasting matters for unit commitment and market bidding. Intraday and ramp forecasting matter for balancing-market exposure and grid stability. The intraday side has been the slower-moving area because it depends on assimilating real-time radar, satellite and sensor data in ways that the older NWP cycle could not. The newer neural models — particularly the diffusion-based GenCast and the rapid-update variants from NVIDIA and DeepMind — are starting to address this, with sub-hourly forecast products now available from several commercial vendors and a growing in-house capability at the larger utilities and aggregators.

ISO and market integration#

The forecasting product only matters when it can be ingested into the bidding and dispatch loops that determine commercial outcomes. CAISO, ERCOT, MISO and the European day-ahead markets have all evolved their gate-closure and intraday-update mechanics to accommodate higher renewable penetration, and the better operators run automated bidding loops that re-optimise positions multiple times per day as forecasts update. The data engineering work — pulling ISO LMP feeds, vendor forecast feeds and plant SCADA into a single decision loop — is non-trivial and is where in-house teams add most of their value beyond what off-the-shelf forecast vendors deliver.

Where the technology still struggles#

Three weak spots remain. Extreme-weather forecasting — the 2021 Texas freeze, atmospheric rivers in California, heat domes — is where neural NWP models still have meaningful gaps versus physics-based models, since the training data contains few extreme events. Offshore wind in waters with poor historical sensor coverage suffers from the same. And the rapid-update intraday side is still less mature than the day-ahead side. The honest 2026 picture is that neural NWP has lifted day-ahead accuracy meaningfully but extreme-event and intraday gaps remain.

Where pdpspectra fits#

Our ML and MLOps practice helps renewable operators and aggregators build the forecast-blending, asset-level post-processing and automated bidding loops that turn raw vendor forecasts into commercial outcomes. We work across DNV, Solcast, Vaisala and self-hosted neural NWP pipelines.

Related reading: Australia renewables grid in 2026, UK energy grid renewables in 2026, and Brazil energy grid AI on ONS data.

---

Renewable forecasting in 2026 is in the middle of a neural-NWP transition that is still playing out. Talk to our team about your forecasting stack.