Abstract
This work investigates the benefits of an incremental training strategy designed to emulate realistic operational updates, and the impact of using complete irradiance inputs — including direct normal, diffuse horizontal, and global horizontal irradiance — for bifacial photovoltaic power prediction. Using data from a Nordic climate setup with four bifacial photovoltaic modules equipped with dual-axis sun trackers, three classical machine learning models and one long short-term memory model were trained to predict instantaneous and short-horizon power output. The proposed incremental strategy was found to be sensitive to training-month order; to address this, multiple runs with randomized month-order sequences were performed. Zero-shot predictions, corresponding to months not used in training, produced median coefficient-of-determination values ranging from 0.380 to 0.964 depending on model and module orientation. Models trained with global horizontal irradiance alone exhibited lower accuracy and greater sensitivity to training data selection. Feature-importance analysis further showed that direct normal and diffuse horizontal irradiance contributed substantially to prediction accuracy, particularly under low-sun and snow-affected conditions typical of Nordic environments.