Accurate power models for photovoltaics (PV) are essential to ensure expected system performance, thus avoiding economic losses and reductions in resource efficiency. Currently, insights are lacking on how reliably models perform in situations for which historical data are not available and how accurately they cover locations with high seasonality, such as the Nordics, where PV capacity is growing rapidly. Importantly, besides long summer days, in winter months, the Nordic areas experience low irradiances, for which power modelling is poorly understood but should be expanded to enable continuous monitoring. To fill current gaps in the literature, this work compares various empirical models (Huld, PVWatts, PVUSA) and investigates how significant improvements can be made with machine learning (ML) models (multilayer perceptron, gradient boosting). The models are analysed as general models, which can be applied directly to new systems, and as site-specific fine-tuned models, for which previous data from that system are required. The data include multi-year power output and on-site weather measurements from five systems across Finland with varying installations. Novel insights include that utilising transfer learning with high-quality data resulted in R² values of 0.944–0.994, and superior accuracy compared to the default models with minimal filtering that had R² values even as low as 0.78. Fine-tuning with site-specific data increased the R² values to 0.966–0.997. Additionally, fine-tuned ML models had lower nRMSE and MAE values than empirical models at low irradiance levels, highlighting their potential for low-light monitoring.