Performance Loss Rate Consistency and Uncertainty Across Multiple Methods and Filtering Criteria

As part of our IEA-PVPS Task 13 research on performance loss rate (PLR) determination we report a comparison between several common approaches, including the effects of different data filtering criteria and power predictive models, resulting in 20 distinct PLR calculations. The open data used for the comparisons is from the DOE Regional Test Center Baseline project consisting of 8 identical PV Systems in 4 distinct Köppen-Geiger climate zones, with 1-minute resolution power and weather data. Five filtering methods were compared: three irradiance cutoffs of 0, 200, 800W/m
²
POA, along with two clear sky filters using a clear sky index, and a moving average detection method. Four different power predictive models of varying complexity were compared: the 6K model, the PVUSA model, the data-driven XbX model, and the XbX + UTC model which is comparable to a temperature corrected performance ratio. PLR uncertainty is evaluated through bootstrap iteration of daily predicted power results. PLR comparisons show the XbX + UTC model has the most consistent PLR magnitude across different filtering techniques, with 6K predicting PLR values inconsistent with the other methods in all locations except Florida's Cfa climate zone. System performance by climate zone showed the highest power loss was observed in Florida (Cfa, -1.84%/year), followed by Nevada (BWh, -0.776%/year), then New Mexico (BSk, -0.54%/year), and lastly Vermont (Dfb) showing the best performance (Dfb, 0.063%/year). The 6K model uncertainty is lowest when retaining as much data as possible, and XbX and PVUSA have the highest stability with high irradiance cutoffs, typically used to reduce the temperature variation of the modules. XbX + UTC showed the lowest PLR uncertainty in each filter method. In all cases, the moving average clear sky filter led to the highest PLR uncertainty. Large fluctuations in calculated PLR suggest PLR results are strongly biased by the filtering and method chosen and more validation is needed for individual models.