A solar forecast fix cut percentage error—mostly by learning when output is zero
ZAPP reduced a volatile percentage-error measure across six forecasting approaches at two Turkish solar plants. The correction mainly removed small night-time forecasts; overall power-scale errors barely changed.
The short version
3 key points
- What changed
- The study used 15-minute power and weather records from two 1 MW solar plants in Türkiye, covering 1 August 2022 to 3 August 2024; the exact retained record counts were not reported.
- Why it matters
- More reliable short-term forecasts can help operators understand expected solar output, but this experiment did not test whether the correction reduced balancing costs, outages or consumer bills.
- What to watch
- Locked multisite tests in other climates, exact record denominators, longer horizons, probabilistic uncertainty, seasonal drift checks and evidence that zero-aware forecasts improve real grid or plant decisions.
The system fixes a metric problem that operators can still care about
Solar forecasts often produce tiny positive values at night. Those misses barely affect errors measured in kilowatts, but percentage errors can explode when the true output is zero. ZAPP adds two steps after a base forecast: a classifier decides whether the next output should be exactly zero, then a constrained rule trims remaining low-power predictions. It does not replace the underlying forecasting model.[1]
Who it affects
More reliable short-term forecasts can help operators understand expected solar output, but this experiment did not test whether the correction reduced balancing costs, outages or consumer bills.[1]
Keep in mind
Two similar 1 MW Turkish plants; exact retained record counts absent; one 15-minute horizon; deterministic forecasts; validation-tuned rules; no operational, cost, reliability or emissions outcome.
Source analysis · 1 linked source. Original source date: 17 October 2026.
Continue the story
Read next
Climate & Energy
A grid anomaly model scored 95% accuracy—but caught none of the faults
A transformer fitted renewable-energy time series well and reached 0.89 ROC-AUC, yet its default threshold missed all 10 labelled anomalies. The result shows why operators need class-specific recall, not headline accuracy alone.
2 min · 1 source
Climate & Energy
Can machine learning make two-to-six-week weather forecasts useful?
A Nature study found that probabilistic bias correction improved leading dynamical and AI systems and won ECMWF’s 2025 real-time competition. Better forecast scores are promising, but benefits for farms, grids and emergency response have not yet been measured.
2 min · 2 sources
Read more The full report, methods and context
The system fixes a metric problem that operators can still care about
Solar forecasts often produce tiny positive values at night. Those misses barely affect errors measured in kilowatts, but percentage errors can explode when the true output is zero. ZAPP adds two steps after a base forecast: a classifier decides whether the next output should be exactly zero, then a constrained rule trims remaining low-power predictions. It does not replace the underlying forecasting model.
The researchers tested persistence, Ridge, XGBoost, LSTM, Transformer and a Ridge–XGBoost hybrid on two 1 MW Turkish plants. Records covered two years at 15-minute intervals; 72 hours of history predicted the next 15 minutes. The first 72% of each time series trained models, 8% tuned them and the final 20% tested them. The paper does not state the exact number of retained records after duplicate handling and gap filling.[1]
The dramatic percentage gain needs the power-scale numbers beside it
For the hybrid at plant one, symmetric mean absolute percentage error fell from 60.57% to 11.07%. At plant two it fell from 63.30% to 11.92%. But plant-one mean absolute error moved from 22.56 to 22.37 and RMSE from 60.72 to 60.72; R² stayed about 0.964. The post-processing mainly corrected zero and near-zero periods rather than changing the overall production curve.
That distinction matters for grid teams. A cleaner zero-output forecast may improve dashboards and relative-error reporting, while dispatch and reserve decisions depend on absolute errors, forecast horizons and uncertainty during meaningful generation. The study tested deterministic one-step forecasts only; it did not measure operating costs, balancing actions, curtailment, reliability or emissions.[1]
The distinction that matters
Where the large improvement came from
The post-processing changed zero and low-output forecasts while leaving the daytime power profile largely intact.
Base forecast
Six models predict the next 15-minute output
Good overall R², but small positive night-time estimates
ZeroGate
Classify likely zero-output periods
Replace selected forecasts with zero
Low-power rule
Trim remaining small forecasts under validation constraints
Large sMAPE reduction; little change in MAE or RMSE
Two sites are replication, not global validation
Both datasets came from 1 MW plants in Türkiye and used the same 15-minute horizon and two-year window. Chronological splitting reduces future-data leakage, code and data are linked, and the method worked across six forecasting families. Still, the threshold and low-power shrinkage were tuned on each validation period, and the paper warns that cloudy winter conditions could make genuine low output look like zero.
The case would strengthen through locked tests across different climates, plant capacities, inverter types, seasons and forecast horizons, with probabilistic uncertainty and operational outcomes reported. Until then, ZAPP is a useful metric-aware correction for two sites—not evidence that solar forecasting broadly became five times more accurate.[1]
Evidence check
Two similar 1 MW Turkish plants; exact retained record counts absent; one 15-minute horizon; deterministic forecasts; validation-tuned rules; no operational, cost, reliability or emissions outcome.
Reporting basis: Source analysis. Original source date: 17 October 2026.
Explore the full source and evidence record
What this means for people
- More reliable short-term forecasts can help operators understand expected solar output, but this experiment did not test whether the correction reduced balancing costs, outages or consumer bills.
Global context
Türkiye spans contrasting solar conditions, but two plants cannot represent the climates, hardware and market rules faced by solar operators worldwide. Near-zero behaviour also differs for wind and other renewable sources.
What the evidence does not yet show
- Two similar 1 MW Turkish plants; exact retained record counts absent; one 15-minute horizon; deterministic forecasts; validation-tuned rules; no operational, cost, reliability or emissions outcome.
What to watch next
- Locked multisite tests in other climates, exact record denominators, longer horizons, probabilistic uncertainty, seasonal drift checks and evidence that zero-aware forecasts improve real grid or plant decisions.
Living evidence record
Impact record IAI-1L5Y1NU
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
18 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Scientific Reports published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 18 October 2026
Research question, topics and regions
Whether explicit zero-output detection and constrained low-power calibration reduce misleading percentage errors in short-term photovoltaic forecasts
Topics: Solar power · Forecasting · Electricity grids · Machine learning · Evaluation metrics
Regions: Türkiye · Europe · Asia · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.