support

This comprehensive documentation outlines the logic of the sensor-health KPI designed to select the best performing irradiance sensors for each inverter of a PV site on a daily basis. Figure 1 shows commonly used sensor types for measuring solar irradiance for PV plants. On the left side of the image is a reference cell, which usually uses the same PV technology as the one installed in the plant, and on the right a pyranometer.

The purpose of this document is to provide insights about the methodology employed, providing confidence in Performance+'s reliability and accuracy. The document covers the entire high-level process, from data collection to the final selection of the individual sensors and the preparation of the final sensor table.

Chapter 1 outlines the high-level approach of how the sensor-health KPI is used to determine the most reliable irradiance sensor, while Chapter 2 provides details of how the sensor-health KPI is calculated. Chapter 3 closes with how the individual irradiance sensors are assigned to the inverter.

Figure 1: Typical solar irradiance sensors

1. High-level architecture

The sensor-health KPI (SH KPI) is a key performance indicator used to evaluate and compare individual sensors with one another. The aim is to decide which sensor is the most reliable for each day for each measurement type. The SH KPI rates each individual sensor based on individual factors in order to make an informed decision about which of the available sensors is most suited to be used for all downstream calculations. Figure 2 shows an example of a PV site and 4 installed plane-of-array (POA) sensors to choose from.

Figure 2: Example of plane-of-array sensor data for one day of one PV site

The SH KPI runs from 0-100. Increasing values represent higher quality readings of the sensors. A SH KPI of -99 represents a disregarded sensor which did not pass one or several of the data reliability tests. The SH KPI is an aggregate across several data quality and reliability related parameters:

Figure 3: Sensor-health KPI structure

The sensor-health KPI is calculated for every ground-based sensor in order to select the best suited sensor per inverter for downstream calculation. The SH KPI is calculated on a daily basis, and the sensor selection per inverter also runs on a daily basis.

2. Sensor Health KPI

As introduced in Figure 3, the SH KPI is constructed based on several parameters. In section 2.1, the individual parameters are introduced. In section 2.2, the calculation of the SH KPI is described. Chapter 3 explains how it is used to select the appropriate irradiance sensor per inverter.

2.1 Parameters

2.1.1 Data quality

The first set of parameters are data quality (DQ) parameters. The following DQ parameters are considered:

• Missing data: NaN values
• Corrupt data: corrupt values are commonly one of the following: [-99, -999, -9999, ... ]
• Stuck data: stuck or unchanging values are identified by checking whether data points vary as expected over time
• Out-of-bounds data: values outside acceptable measurement ranges

Every datapoint during daytime (filtered based on elevation data) is subject to data quality tests. If more than 40% of daytime datapoints are affected by DQ issues, the sensor is hard-excluded for that day (SH KPI = -99). For sensors below that limit, the DQ parameter used in the score is the fraction of daytime timestamps flagged as bad: the lower the fraction, the higher the score - see Table 1 for the exact bands. Data quality is evaluated in an earlier stage of the Performance+ pipeline; the results feed directly into the sensor-health KPI as described here. Figure 4 shows how these data quality flags are presented in BazeField.

Figure 4: Data quality overview of irradiance sensors (to be replaced with native front-end once available)

2.1.2 DC power correlation parameter

Power and irradiance are highly (almost linearly) correlated. That is why this correlation parameter was introduced and carries a high weight in the determination of the SH KPI. The objective is to select one representative DC input (MPPT or combiner) per racking group, so that its plane-of-array orientation (tilt & azimuth) is aligned with the orientation of the irradiance sensor. Because the sensor-health KPI runs early in the pipeline, the DC power sources are only subject to basic data quality filtering (curtailment removal and daytime filtering), not yet to the extended reliability/IEC checks applied later in the pipeline.

The parameter is prepared per racking group, per day, as follows:

1. Map and filter. Each inverter is mapped to its racking group from the asset metadata, and its AC power is filtered to non-curtailed daytime periods.
2. Best inverter. For each day, the capacity-normalized daily AC yield is computed for every inverter in the racking group, and the highest-yielding inverter is selected.
3. Best DC input. For that highest-yielding inverter, each of its DC inputs (MPPT/combiner) is normalized by its capacity, and the DC input with the highest normalized daily yield is selected. The full-resolution series of these per-day winners is stitched together into one continuous DC-power signal per racking group, which is then correlated against the irradiance sensor readings.

This yields a single, orientation-aligned DC input per racking group while always tracking the best-performing equipment on each day. The parameter value used in the SH KPI is the daily Pearson correlation between the normalized DC power and the irradiance sensor reading; the score bands are listed in Table 1.

2.1.3 Satellite correlation parameter

The satellite correlation parameter is the daily Pearson correlation between the sensor reading and the corresponding satellite reanalysis - GHI sensors against raw satellite GHI, POA sensors against satellite GHI transposed to the sensor plane. The resulting correlation feeds directly into the SH KPI score; the bands are listed in Table 1.

2.1.4 Satellite bias

In addition to the satellite correlation (section 2.1.3), the absolute daily agreement with satellite reanalysis is evaluated as a bias check. The bias is the daily daytime mean absolute difference |sensor - satellite|. For irradiance sensors it is used only for removal - it does not contribute to the graded SH KPI score. A gross bias triggers an immediate hard flag (SH KPI = -99) that discards the sensor for that day, independent of all other parameters. The default gross-bias threshold for irradiance is 150 W/m^2.

2.1.5 Transposed clear-sky model correlation parameter

The clear-sky (CS) model is modelled based on the Ineichen and Perez clear-sky model [1, 2] available in the Python package pvlib [3]. To transform the global horizontal clear-sky model into POA, decomposition and transposition models are used based on [4, 5, 6]. The parameter is only used for the SH KPI on clear-sky days. A clear-sky day is defined as a day in which at least 80% of the datapoints are considered to be clear-sky instances. The parameter value used in the SH KPI is the daily Pearson correlation between the modelled clear-sky irradiance and the sensor reading; the score bands are listed in Table 1.

2.1.6 Irradiance deviation

Irradiance deviation is based on the reference irradiance introduced in section 2.1.3. Irradiance deviation is only checked on clear-sky days. At each timestamp, the median value of all POA sensors with the same specified orientation is calculated. Then, for each sensor, the percentage error between the total daily irradiance measured and the median value is determined:

POA_error = (sum POA_meas - sum Irr_ref) / sum Irr_ref

The sum is always daily. Here, POA_meas are the sensor readings and Irr_ref is the reference irradiance. The absolute value of this daily deviation is the parameter used in the SH KPI score; a deviation below 5% scores full marks, between 5% and 10% scores medium, and above 10% scores minimum. If the deviation exceeds 12%, the sensor is hard-excluded for that day regardless of all other parameters - see Table 1 and the hard-exclusion table for the full details.

2.2 Sensor Health KPI calculation

The SH KPI is an aggregated KPI based on the results from the quality test parameters described above. Table 1 describes all parameters, their fail thresholds, as well as their score levels to calculate the final SH KPI. The KPI is calculated based on the following formula:

Sensor-Health KPI = (sum score x 100) / sum max_score   [%]

Here, max_score is the maximum reachable sum if all parameters under consideration return the highest possible score. score is the sum of all individual scores of the parameters considered.

Table 1: Graded sensor-health KPI parameters (irradiance sensors)

^1 Power-irradiance correlation carries double weight: only the top band contributes 10/10; the lower bands contribute out of 5, so the denominator adapts to the band reached. ^2 On non-clear-sky days the irradiance-deviation threshold is relaxed by a factor of 4.

If any of the parameters in Table 1 is not available for a given day, the SH KPI is computed excluding that parameter (for example CS-irradiance correlation and irradiance deviation are skipped on non-clear-sky days). GHI sensors are graded without the power-irradiance correlation.

Hard exclusions (SH KPI = -99). Independent of the graded score above, a sensor is discarded for a given day whenever a sensor-check flag is raised:

2.3 Non-irradiance sensors (ambient temperature and wind speed).

Table 1 grades irradiance sensors. Ambient temperature and wind speed sensors are graded on (i) data quality, (ii) own deviation/satellite checks, and (iii) satellite bias (section 2.1.4). The latter has a higher weight (x2) into their SH KPI score: a bias within the threshold (3 degC / 2.5 m/s) scores full marks, twice the threshold scores medium, more scores zero. Ambient temperature also carries a gross-bias hard flag at 10 degC.

2.4 Results

The final SH KPIs across a PV site and for each day can be seen exemplarily in Figure 5. It is visible that global horizontal irradiance (GHI) sensors are also evaluated, but without considering power-irradiance correlation. A SH KPI of -99 marks a sensor disregarded by one of the hard exclusions above.

Figure 5: Example of sensor-health KPI heatmap

3. Irradiance sensor selection

The final step is to pass the SH KPI information to the irradiance source selection. Here, the best performing irradiance source (sensor or satellite) is selected for each day for each inverter. The irradiance selection follows a strict prioritization scheme which can be adjusted based on the requirements of the customer. Table 2 shows the default POA priority order.

Table 2: Default irradiance source prioritization (POA)

GHI sensors follow the same order without step 4 (no transposition needed), falling back directly to satellite GHI. We prioritize POA sensors first, followed by GHI sensors transposed to POA.

We run an initial selection across the ground-based irradiance sources - assigned -> average -> highest-rated POA -> highest-rated transposed GHI - in that priority order. A source is accepted in this first pass if its SH KPI is at least 85%. If none of the ground-based sources reaches 85%, we run the same prioritization again with a lowered threshold of 65%. If no ground-based source reaches at least 65%, we fall back onto satellite data (direct satellite POA, or satellite GHI transposed to the sensor plane), as we then have to assume that all sensors are corrupt or of very low quality. If no satellite data is available, we fall back onto the ground-based source with the highest remaining SH KPI; and if every sensor failed its hard quality tests (SH KPI = -99), we use the source with the most valid datapoints (highest DQ parameter). Figure 6 summarizes this logic.

The further we fall back in our prioritization, the more our downstream analytics results will be affected by non-optimal irradiance input data.

Figure 6: Irradiance sensor selection process

References

[1] R. Perez and P. Ineichen, "A new airmass independent formulation for the Linke turbidity coefficient," Solar Energy, vol. 73, pp. 151-157, 2002.

[2] R. Perez et al., "A new operational model for satellite-derived irradiances: description and validation," Solar Energy, vol. 73, pp. 307-317, 2002.

[3] W. F. Holmgren, C. W. Hansen and M. A. Mikofski, "pvlib python: a python package for modeling solar energy systems," Journal of Open Source Software, vol. 3, no. 29, 2018.

[4] R. Perez, P. Ineichen, E. Maxwell, R. Seals and A. Zelenka, "Dynamic Global-to-Direct Irradiance Conversion Models," ASHRAE Transactions-Research Series, pp. 354-369, 1992.

[5] R. Perez, R. Seals, P. Ineichen, R. Stewart and D. Menicucci, "A new simplified version of the Perez diffuse irradiance model for tilted surfaces," Solar Energy, vol. 39, no. 3, pp. 221-232.

[6] M. Mikofski and K. Anderson, "Slope-Aware Backtracking for Single-Axis Trackers," NREL, 2020.