Abstract
This study investigates the applicability and benefits of day-ahead probabilistic photovoltaic power forecasting at high latitudes by coupling ensemble numerical weather predictions (NWP) with a hybrid physical-statistical modeling framework. The analysis addresses typical challenges of high-latitude regions, including low solar elevation, frequent overcast conditions, and snow accumulation on photovoltaic modules. A high-resolution dataset from a building-applied photovoltaic system in Trondheim, Norway, is used to analyze predictive performance across multiple meteorological regimes and seasons. Experimental validation shows that NWP provides reliable inputs for probabilistic photovoltaic forecasting, with strong discriminative skill for air temperature and solar irradiance (ROC ≃ 0.9) and moderate skill for wind speed and snow depth (ROC ≃ 0.7), reflecting unresolved local-scale
processes. When considering photovoltaic power forecasts, raw ensemble-based physical model chains deliver reasonable probabilistic forecasts but suffer from systematic biases and structural under-dispersion, with nMBE exceeding 5%. The application of rolling-window statistical calibration improves forecast quality, reducing nMBE to 2%, while lowering normalized mean absolute error by approximately 2%. In probabilistic terms, calibration enhances uncertainty representation, reducing the CRPS-to-spread ratio from values above 1.5 to values close to unity for predicted photovoltaic power. The error propagation analysis reveals that irradiance decomposition and transposition are the main sources of uncertainty amplification, particularly under overcast conditions, whereas the irradiance-to-power conversion contributes less to the overall bias. In conclusion, the results demonstrate that
hybrid probabilistic approaches can provide reliable photovoltaic power forecasts at high latitudes, supporting robust solar integration in cold climates and remote communities.