Photovoltaic system monitoring for high latitude locations

•Snow, irradiance differences and low light conditions reduce monitoring reliability.•Tailored filters reduce variability and keep more data compared to standard filters.•Machine learning is efficient for PV monitoring in locations with variable climate.•Topography and shading lead to systematic irradiance differences in PV systems.•Unstable PV low light behavior is related to irradiance level and solar position. Reliable monitoring of PV systems is essential to establish efficient maintenance routines that minimize the levelized cost of electricity. The existing solutions for affordable monitoring of commercial PV systems are however inadequate for climates where snow and highly varying weather result in unstable performance metrics. The aim of this work is to decrease this instability to enable more reliable monitoring solutions for PV systems installed in these climates. Different performance metrics have been tested on Norwegian installations with a total installed capacity of 3.3 MW: (i) comparison of specific yield, (ii) temperature corrected performance ratio, and (iii) power performance index based on both physical modelling and machine learning. The most influential effects leading to instability are identified as snow, low light, curtailment, and systematic irradiance differences over the system. The standard deviation of all the performance metrics is reduced when filters targeting these four effects are applied. Compared to general low irradiance or clear sky filtering, a greater reduction in the variation of the metrics is achieved, and more data remains in the useful dataset. The most suitable performance metrics are comparison of specific yield and performance index based on machine learning modelling. The analysis highlights two paths to accomplish increased reliability of PV monitoring systems without increased hardware costs. First, better reliability can be achieved by selecting a suitable performance metric. Second, the variability of the performance metric can be reduced by utilizing filters that specifically target the origin of the variability instead of using standard literature thresholds.