Abstract
This study investigates the potential of machine learning (ML) tools to enhance sub-hourly photovoltaic (PV) power simulations in high-latitude locations with snowy conditions. A comparative analysis was conducted between calibrated physical PV model chains and ML tools to assess their ability to estimate snow losses (i.e., PV power output reduction due to snow accumulation). An ensemble of 100 physical PV model chains was implemented to probabilistically estimate the generated power. Additionally, five ML methods – gradient boosting regressor, linear regression, decision tree regressor, random forest regressor, and multi-layer perceptron regressor – were used both to i) calibrate the physical model ensembles (PME) and to ii) directly simulate PV power output. Model training, calibration, and validation were performed using observational data from the PV modules installed at the ZEB Test Cell Laboratory in Trondheim, Norway. PV power output was monitored with a 15-minute resolution from November 2023 to April 2024. The findings demonstrated that ML models, particularly random forest regressor, could outperform PME in capturing complex meteorological interactions such as snow accumulation, melting, and irradiance variability. This highlights the potential of ML tools to reduce uncertainties in solar analyses and support the design of more resilient and adaptive PV systems for the Subarctic and Arctic regions.