Short-Term PV Forecasting Using IDW-Based Weather Estimation and a Comparative Analysis of XGBoost Loss Functions

High PV penetration increases day-ahead scheduling risk by amplifying peak and ramp errors. We examine how spatial interpolation strength and loss-function choice affect short-term PV forecasts. Using Jeju Island data (plant outputs, 2020-01 – 2024-05; ASOS stations Jeju and Gosan), we estimate plant-level meteorology via inverse distance weighting with exponent p∈{1,…,5} plus a spatial-average baseline, and train XGBoost models under three regression losses like MSE, MAE, and Pseudo-Huber. Training is restricted to daylight hours, and performance is evaluated using normalized mean absolute error. Results reveal strong seasonal interactions between interpolation strength and loss: the best seasonal nMAE is 5.48% (spring, IDW p=4 and MAE), 7.12% (summer, Average and Pseudo-Huber), 6.32% (autumn, IDW p=1 and Pseudo-Huber), and 5.42% (winter, IDW p=4 and MAE). These findings indicate that season-aware configuration of IDW and robust losses enhances peak/ramp tracking and reduces nMAE under sparse weather networks.