Scientists created a physics‑informed multimodal deep learning framework called IMPACT for probabilistic photovoltaic power forecasting between 10 and 110 minutes ahead. The design uses four separate encoders: a plant‑level temporal history branch, an inverter‑level spatial cloud dynamics branch with a graph‑attention encoder, a horizon‑covariate branch containing solar geometry and ramp‑down activity, and a meteorological snapshot branch. Gated cross‑attention fuses the four streams, and a decoder anchors predictions to a smart persistence baseline, applying an online kernel‑weighted adaptive conformal calibration. Across three expanding‑window cross‑validation splits on data from a 205‑MW solar plant, IMPACT outperformed 14 baseline models, achieving a forecast skill score (FSS) of 0.2950, at least 0.099 higher than the next best model. Its deterministic errors were the lowest among competitors: normalized root‑mean‑square error (nRMSE) 0.1375, normalized mean absolute error (nMAE) 0.0669, and R² of 0.7683. The model’s performance was lower under overcast sky conditions, with the lowest forecast skill (FSS = 0.198) compared to clear skies.