In utility-scale solar photovoltaic (PV) power plants, accurate long-term energy yield assessment is contingent on the continuous, uninterrupted availability of high-quality meteorological data. Instrumentation faults, sensor drift, and transient measurement anomalies if left undetected propagate directly into performance ratio (PR) calculations, irradiance loss estimates, and contractual energy yield guarantees, introducing systematic errors that are difficult to isolate and correct retrospectively.
To mitigate these risks, modern utility-scale PV projects have widely adopted the practice of deploying redundant sensor pairs within each Weather Monitoring Station (WMS). This architecture satisfies the redundancy requirements of IEC 61724-1:2021 and simultaneously enables the operations and maintenance (OCM) team to perform continuous cross- validation of sensor outputs, thereby ensuring the sustained measurement integrity of the WMS over the project lifetime.
However, a recurring operational challenge has been observed in the field: onsite OCM and project engineering teams frequently apply inappropriate comparison methods when evaluating the data from redundant sensor pairs. The most common malpractice is the direct per-minute comparison of instantaneous readings from two co-located sensors. Under this approach, legitimate atmospheric phenomena including cloud transients, thermal gradients, and turbulent wind events produce large instantaneous deviations between the two channels, causing sensors to be incorrectly flagged as faulty. This misdiagnosis results in unnecessary field interventions, unjustified sensor replacements and critically, a loss of confidence in instrumentation that is in fact performing within specification.
This paper presents a statistically rigorous and practically validated methodology for the quality control (ǪC) of redundant WMS sensor pairs at utility-scale solar PV plants. The proposed framework, implemented and verified across a GW utility scale Solar PV Power plant in the Middle East, applies five complementary evaluation methods - diurnal trend matching, Pearson correlation (r) analysis, maximum and minimum absolute deviation assessment, Drift Index ranking, and multi-day trend evaluation - to distinguish genuine sensor faults from atmospheric variability with a high degree of diagnostic confidence. Results covering a 10-day continuous monitoring window (02–12 November 2025) across seven stations and five parameter groups are presented, demonstrating the efficacy of the methodology under real field operating conditions.
Often, solar plant operators, EPCs, and OCM teams raise the same alarm: "Your two sensors are showing different values at a specific time. One of them must be faulty.
In almost every case, both sensors are working perfectly. The problem isn't the instruments it's the method of comparison.
This report is a definitive technical guide for anyone comparing redundant WMS sensors on a utility-scale solar site. We will show you - with real data and statistics why per-minute comparison is fundamentally flawed, and why statistical validation using daily averages, standard deviation, and correlation coefficients is the correct and IEC-aligned approach.
Key Insights : Two Class A sensors installed side by side on the same mast will rarely show identical per-minute readings. The atmosphere is not uniform at millisecond timescales. This is physics - not instrument failure.
When two sensors are compared at every 1-minute interval, instantaneous atmospheric variability, turbulence cells, shadow transients, and sensor response lag all contribute to apparent "disagreement" even when both sensors are perfectly calibrated and within Class A tolerance.
For example, a single per-minute deviation of 500 W/m² during a cloud transient is not a fault it is atmospheric physics. Cloud edges can pass one pyranometer 30–60 seconds before the adjacent sensor, producing a large instantaneous delta while the daily aggregate correlation remains above 0.999. The Drift Index and daily Pearson Correlation (r) correctly distinguish transient weather events from genuine sensor faults.
The data analysis framework applies five complementary statistical methods to validate redundant sensor pairs.
Using any single method in isolation is insufficient; the combination of all five provides a robust, falsification-resistant view of sensor health.
We retrieved a 10-day WMS data from 7 Weather Monitoring Stations (WMS) commissioned at a 2GW Solar PV Power plant project in the Middle East.
4 most critical parameters like Pyranometer (GHI), Albedometer, Module Temperature Sensor and Wind Speed sensor data were thoroughly analysed. The data was recorded at 1 minute interval in the Aeron make XG930 Class A data logger of the WMS.
Figure 1: GHI (Horizontal Pyranometer) clear sky profile
This is the gold standard of what healthy redundant sensors look like: near-perfect Pearson correlation (r), maximum divergence under 2% of peak, and no directional drift across the day.
A large cloud shadow event passed across Station 45 between 12:53 and 13:04. Both GHI sensors responded simultaneously but by different amounts, creating a large instantaneous delta.
Figure 2: Station 45 · 03 Nov 2025 — GHI Transient Event (cloud shadow at 12:53–13:04)
GHI_IRR_01_01 maintained a clean, uninterrupted bell curve throughout the day. GHI_IRR_02_02 showed a single sharp transient dip of −482 W/m² at 10:12, recovering fully within 1 minute. This is the most diagnostically important event in the dataset: sensor 1 proves the true sky irradiance while sensor 2’s isolated dip confirms a single-sensor event and not necessarily a sensor fault (likely bird strike or any other phenomenon).
Figure 3:Station 41 · 09 Nov 2025 — GHI dip (GHI_IRR_02_02 transient at 10:12)
ALM_Up (upward-facing, measures incoming global irradiance) and ALM_Down (downward-facing, measures reflected ground irradiance) are both validated as paired sensors. Station 41 shows textbook albedometer behaviour on clear days
Figure 4: Albedometer (ALM) Analysis
Figure 5: 10 Days ALM correlation
Station 45 (03 Nov) dips to 0.3C48 and Station 50 (03 Nov) to 0.34C8 — both correctly identified as cloud transient events by the Drift Index, not sensor faults.
Module temperature (MTS) follows the sun: it rises with irradiance in the morning, peaks in the early afternoon (lagging solar noon by ~1–2 hours due to thermal mass), then drops as irradiance fades. Two sensors at the same station should trace nearly identical curves all day. Below is Station 54 on 05 Nov - one of the cleanest days in the dataset.
Figure 6: MTS healthy profile
This is the most striking MTS event in the entire 10-day dataset. MTS_02_01 at Station 72 recorded a single reading of ~89°C at 14:18 - a jump of roughly 35°C above the other sensors at that moment. All other sensors at the same station continued their normal afternoon decline. The blue sensor stays smooth; only the red sensor spikes.
Figure 7: Sudden spike in MTS02_Stn. 72
A detailed analysis revealed that, the MTS_02_01 had MaxDiff = 89.35°C @ 14:18 while all other MTS channels at the same station showed MaxDiff < 3°C for the same period.
Corr (r)= 0.972 - the only instance in the entire 10-day dataset where any MTS sensor fell below (r) = 0.980. The Drift Index ranked this as the only MTS anomaly across all 7 stations and all 10 days.
Likely cause: loose thermocouple connection causing a momentary open-circuit reading, or localised soiling/hotspot on that specific cell.
Figure 8: 10 Days MTS correlation
The view shows each station as a time-series line. You can immediately see that 6 out of 7 stations hold R ≥ 0.998 across all 10 days without exception. Station 72 on 06 Nov is the single outlier and it stands out clearly as a dip isolated to one day on one station.
Unlike GHI or MTS, wind speed is naturally turbulent it fluctuates second-to-second due to eddies, terrain effects, and mast shadow. Two anemometers at the same station will never trace identical curves the way pyranometers do. A healthy day looks like two noisy but statistically similar signals, tracking the same general gusts and lulls. Below is Station 41 on 02 Nov a good reference day with R = 0.964.
Figure 9: Wind Speed sensor healthy profile
Figure 10: Wind Speed sensor WS_02_02 fault identified
This is a genuine sensor fault, not a weather event.
WS_01_01 (blue) shows a normal breezy afternoon with gusts up to ~6 m/s while WS_02_02 (red) starts normally in the morning, then from ~10:30 onwards produces erratic, physically implausible readings - including a MinDiff of −24.49 m/s, which is impossible for an anemometer. The two signals become completely uncorrelated.
Figure 11: 10 Days Wind Speed sensor correlation
Wind speed naturally produces more day-to-day variation in (r) than GHI or MTS - calm days with little wind push correlation down legitimately. The chart above shows this natural variability across all stations, with the 0.95 threshold marked.
Values dipping to 0.91–0.96 on calm days are normal - low wind means small absolute values and any tiny offset dominates the corelation factor (r). What is not normal is a value near zero (0.0954). The dashed red threshold line at R = 0.95 shows which station-days required investigation. Station 41 on 04 Nov is the only genuine fault in the entire dataset - all other dips are within expected atmospheric variability.
A thorough five-method framework used here caught every real anomaly while producing zero false alarms on the 7-station, 10-day dataset.
Real-world WMS data from a running Solar PV site confirms that Pearson correlation (r), daily aggregate deviation, and Drift Index ranking are the correct tools for sensor data analysis.
Per-minute divergence is normal physics. Atmospheric turbulence, cloud transients, and sensor response time all cause instantaneous disagreement between co-located sensors. It should thus not be considered as a sensor fault.
Auther : Saurabh Dey
