DYNVOLT shows 5.1% day-ahead solar forecast error at EU PVSEC 2026

DYNVOLT said its new AI forecasting model hit 5.1% normalized mean absolute error in live day-ahead testing on utility-scale solar plants in North Macedonia. The company unveiled the results at EU PVSEC 2026 in Rotterdam and published its methodology and exclusion ledger for replication.

Why it matters: - DYNVOLT is pitching a tighter link between solar plant operations and power market bidding for utility-scale PV and storage owners. - The company says its latest forecasting result could improve day-ahead nominations and reduce imbalance exposure on real plants.

What happened: - Zoran Aleksikj, co-founder of DYNVOLT and CEO of Dynode.ai, presented the platform at EU PVSEC 2026 in Rotterdam from Sept. 14 to 18. - DYNVOLT exhibited in the Startup Pavilion and Aleksikj pitched on the Startup Solar Launch Pad stage. - The presentation centered on DYNVOLT forecasting model v2, which has been live since September 2026. - DYNVOLT operates in production at utility-scale plants in North Macedonia. - The deployed system monitors more than 840 PV strings at 30-second intervals and uses the forecasts to nominate production on the power exchange.

The details: - DYNVOLT combines solar SCADA, energy management, forecasting, day-ahead and intraday bid construction, battery dispatch optimization, and O&M reporting in one data model. - The platform runs on a secured Linux-based edge server installed on site. - Plant data is buffered through outages and synchronized to the cloud for forecasting, dispatch, and reporting. - DYNVOLT says it does not place bids or take custody of trading revenue. - Export schedules are prepared for the owner or the balance responsible party to submit. - In a walk-forward out-of-sample evaluation from April to August 2026, the model reached 5.1% normalized mean absolute error on daytime day-ahead forecasts, normalized to AC capacity. - The reported nMAE range was 5.0% to 5.2%, and normalized RMSE was 8.2% to 8.7%. - DYNVOLT said the best comparable published academic result, converted to the same convention, is about 8.6% nMAE. - Smart persistence on DYNVOLT’s own plants scored 12.8%. - The model blends 10 numerical weather prediction sources at their true day-ahead issue time. - The system applies per-plant machine learning correction trained on each site’s history and retrains nightly. - Because telemetry, curtailment commands, and market schedules share one data model, the training pipeline automatically excludes hours when a plant was ordered not to produce. - DYNVOLT said a standalone forecast vendor usually cannot make that correction because it does not see the SCADA command log. - P10 to P90 bands are recalibrated daily and achieved 80% empirical coverage. - DYNVOLT publishes its full methodology and exclusion ledger, including each day removed for curtailment or telemetry outage, so the results can be reproduced. - DYNVOLT said the figures were measured against the meter and used to settle real imbalance positions. - The platform is now offered to asset owners, independent power producers and traders across Europe. - DYNVOLT is based in Kavadarci, North Macedonia. - More information is available on the company's website and in the original release.

Between the lines: - DYNVOLT is framing forecast quality as a data-governance problem as much as a modeling problem. - The company’s emphasis on published exclusions and live-plant validation is meant to make the result easier to audit than typical vendor forecast claims. - The one-data-model approach also gives DYNVOLT a built-in advantage when curtailment and market scheduling affect measured output.

What's next: - DYNVOLT is positioning the platform for broader sales to European asset owners, IPPs and traders. - The company is continuing to publish forecast error, worst-day performance and product gaps as it expands.

The bottom line: - DYNVOLT used EU PVSEC 2026 to argue that cleaner plant data and tighter operational integration can materially improve day-ahead solar forecasting.