<![CDATA[Short-term photovoltaic forecasting supports operational planning only when point accuracy, predictive uncertainty, and transferability are evaluated together. This study investigates these requirements using the original Solar Power Generation Data repository, containing fifteen-minute records from two Indian photovoltaic plants over 15 May–17 June 2020. A data audit identifies incomplete inverter panels and a tenfold discrepancy in the recorded DC-to-AC ratio. Consequently, forecasts target normalized aggregate AC power exclusively at timestamps with all 22 inverter records. Histogram-based quantile gradient boosting is evaluated at 15-, 30-, and 60-minute horizons under a fixed chronological fit–validation–calibration–test design. Local models, bidirectional transfer without target-plant model fitting, source-calibrated intervals, and target-recalibrated intervals are distinguished explicitly. Across the six transfer experiments, the rearranged median forecast reduces mean absolute error by 5.4–32.3% relative to ordinary persistence, although ridge regression outperforms it in several comparisons. Uncalibrated nominal 90% transfer intervals attain only 79.6–87.0% empirical coverage. Source calibration raises coverage to 85.1–91.7%, while three days of target calibration yield 88.1–94.2%. Increased coverage can require materially wider intervals and does not consistently improve weighted interval scores. Paired day-block bootstrap intervals reveal that two transfer skill estimates remain compatible with no improvement over persistence. The evidence supports a practical distinction between transferring a forecasting function and transferring its uncertainty calibration. For industrial energy planning, the resulting framework offers a transparent assessment of forecast risk, while the short observation period, missing-record selection, and limited plant metadata preclude claims of universal transferability or realized economic savings.]]>
