Comparison of day-ahead photovoltaic power forecasting models using site-relative NWP deviation features

Accurate day-ahead forecasting of photovoltaic power is essential for grid scheduling, but forecast skill is limited by meteorological uncertainty and by the mismatch between coarse numerical weather prediction (NWP) outputs and site-level conditions. This study develops a framework that combines physically based capacity-factor analysis with data-driven day-ahead forecasting on the Photovoltaic Power Output Dataset, a multi-site dataset. We first use a theoretical capacity factor model to quantify the gap between ideal and measured generation. We then evaluate the main forecasting models used in this study and analyze their behavior under different weather conditions. To better represent local departures from coarse NWP inputs, we construct site-relative deviation features for temperature and irradiance from each station's difference from the contemporaneous multi-site mean. These variables are treated as lightweight correction features rather than meteorological downscaling. Results show that the Light Gradient Boosting Machine (LightGBM) provides the best overall accuracy for the NWP-enhanced benchmark, while RandomForest is more robust in low-power and outlier-prone periods. The site-relative deviation features produce model-dependent gains: the root mean square error (RMSE) of LightGBM decreases from 0.9852 to 0.9511, and the RMSE of RandomForest decreases from 1.1659 to 1.1544, whereas Extreme Gradient Boosting shows no improvement. The revised findings indicate that simple NWP correction features can be operationally useful, but their benefits are modest and should not be interpreted as a substitute for true NWP downscaling.