Authors: Marketa Hulik Jansova, Katarina Blstak Catlosova, Assa Camara, Tomas Cebecauer, Martin Jakubik, Oliver Osvald | Presented at PVPMC 2026 (Albuquerque) |
12-14 May 2026 |
Overview
Ground-measured solar radiation data is never perfect. Instrumental faults, environmental factors, and installation issues introduce biases that silently distort every downstream result, from solar resource model comparisons to plant performance assessment. This study focuses on three questions:
How large and how frequent are undetected quality control (QC) issues in GHI and GTI ground measurements,
How does automated QC performance vary with data resolution, record length, and sensor redundancy, and
How does residual measurement uncertainty propagate into Energy performance index (EPI) assessment.
The findings show that skipping measurement QC introduces errors of a magnitude comparable to the uncertainty of the full PV yield simulation chain, from input data to interannual variability.
Methodology
The study covers three linked analyses:
Occurrence and impact of QC issues: 509 GHI and 77 GTI datasets were evaluated using Solargis expert QC procedures, quantifying which issue types occur, how often, and their effect on Bias and Mean Absolute Error (MAE) relative to the Solargis satellite-based model when left uncorrected.
Automated QC performance: 114 one-year expert-QC'd datasets with two GHI sensors each were independently modified seven ways: aggregated to four coarser timesteps (5, 10, 15, and 60 minutes), reduced to a single sensor, and shortened to six and three months. Automated flagging precision and recall, and the resulting Bias and MAE reduction, were compared against expert QC.
Propagation into PV performance assessment: One year of 5-minute GHI data was injected with controlled soiling and misalignment issues, then simulated to PVOUT using Solargis Evaluate across three PV configurations (fixed tilt at optimal angle, horizontal north-south tracker, and vertical bifacial single-row) at three climatically distinct sites (Alice Springs, Australia; Munich, Germany; Ouarzazate, Morocco).
Key findings
Occurrence and impact of undetected QC issues
Every dataset analyzed contained at least one flagged quality issue. The median proportion of affected data points was modest (3.2% for GHI, 6.1% for GTI), but reached 18% (GHI) and 15% (GTI) at the 90th percentile (p90), and persistent calibration or misalignment issues corrupted up to 47% of data points at p90. Left uncorrected, these issues amplified Bias by 5.7% (GHI) and 4.6% (GTI) at p90, with similar increases in MAE.
Automated QC performance across resolution, record length, and sensor redundancy
Automation matched expert-level performance under favorable conditions: multi-sensor sites at 1-minute resolution achieved an expert-to-automated performance ratio below 1, meaning automation matched or exceeded expert results. That ratio grew substantially (i.e. performance degraded) with coarser aggregation and reached up to 2 at p90 for single-sensor, short, or coarse-resolution datasets. Data length requirements proved more forgiving than expected, with 6 months performing nearly as well as 1 year, and even 3 months viable on typical datasets.
Automated flagging achieved roughly 50% recall of real issues in 1-minute data with sensor redundancy, yet still delivered comparable overall accuracy, since high-impact anomalies were consistently caught even when low-impact ones were missed.
Propagation of residual uncertainty into PV performance assessment
Uncertainty propagation depends strongly on system design. Vertical bifacial configurations were most sensitive to GHI Bias on annual aggregates, with MAE transfer ratios (PVOUT MAE relative to GHI MAE) that can exceed 1, while the typical range for monofacial fixed and tracker systems was 0.7–1. PVOUT Bias closely mirrored GHI Bias in both magnitude and direction across all configurations. Misalignment impact was largest at the highest-latitude site, and north-south tracker systems were least affected by sensor misalignment. A ±2% GHI Bias translates directly into approximately a ±2% EPI deviation: negative GHI Bias inflates EPI and can mask real underperformance, while positive Bias deflates it.
Conclusions and implications for the PV industry
PV yield uncertainty spans the full chain, from input data and simulation models to interannual variability, typically reaching a few percent at p90. This study shows that skipping measurement QC alone introduces errors of comparable magnitude: assuming a normal distribution (where U{p90} ≈ 1.6 × MAE), residual MAEs translate to p90-level uncertainties that rival the full simulation chain.
Audit your data. Every dataset contains QC issues, and in extreme cases, uncorrected Bias alone can reach the magnitude of the full simulation uncertainty chain.
Invest in your setup. Redundant sensors, high resolution, and 6 or more months of data are prerequisites for reliable automated QC. For high-stakes assessments, complement automation with expert review.
Never skip QC before performance evaluation. An uncorrected issue can make a struggling plant look perfectly healthy on paper.