Forecast accuracy under tropical variability.
Thirty measured days on a 54.4 MW AC plant in Camarines Sur, through the Philippine wet season.
Daily normalised MAE across July 2026
Each point is one day's normalised MAE. Mean for the window is 6.05%.
Context
Calabanga CARE is a 54.4 MW AC / 74.17 MWp plant in Camarines Sur. Rapid cloud movement and sudden irradiance shifts make the Philippine wet season the hardest forecasting window of the year, and July sits in the middle of it.
This was HelioExpect's first international deployment and remains the Southeast Asia operating reference.
What we did
Day-ahead and intraday forecasting at 15-minute resolution, with accuracy tracked daily so performance is visible to the operations team rather than asserted.
Result
Normalised MAE across 30 measured days was 6.05%. The best day came in at 3.14%, the worst at 10.82%. P95 deviation averaged 28.64%.
Accuracy detail
Operational impact
- Forecast performance measured daily and visible to the operations team
- Outlier days attributable to weather rather than unexplained model drift
- Site data compounds into model accuracy for neighbouring geographies
Next steps
Outlier handling on unresolved cloud events is where the remaining error sits. A second independent weather source, blended with model disagreement used as a confidence signal, is the established fix for that failure mode.
Talk to the teamMethodology
Window 1–31 July 2026, 30 days with data at 15-minute resolution.
Accuracy is normalised MAE — absolute error divided by MQmax, the maximum actual generation for that day. This is the backend's `MAPE` field reported under its true definition; a textbook MAPE would read differently on the same data.
No incumbent baseline has been measured on this site, so no improvement figure is claimed.
1 day(s) excluded for data completeness, not for accuracy: 2026-07-30 (no generation recorded — plant offline or feed dead). A day is excluded only when it has no actual data, records no generation, generates under 5% of the site's median day, or returns fewer than 80% of the site's median block count. Poor accuracy on a complete day is never grounds for exclusion.
The Philippines settles through WESM rather than an Indian-style deviation settlement mechanism, so no DSM section applies to this site and none is shown. Accuracy is the whole story here.
- Denominator
- MQmax — maximum actual generation for the day
Run the same review on your portfolio.
Send one month of plant and meter data and we will come back with the breakdown.