Photovoltaic power forecasting: A comprehensive review and taxonomy of forecasting methods

Photovoltaic (PV) power forecasting is a foundational capability for operating power systems with high shares of variable renewable generation. However, the methodological landscape is fragmented across physical-based models, statistical time-series approaches, machine learning (ML), and increasingly diverse hybrid and multi-modal pipelines combining SCADA measurements with numerical weather prediction (NWP) and remote-sensing data. This paper provides a unified, practice-oriented synthesis of PV power forecasting methods, linking forecasting tasks and horizons to data requirements, modeling assumptions, and deployment constraints. We first summarize PV plant architecture, the energy conversion chain and key definitional choices (AC vs. DC, power vs. energy, spatial aggregation, sampling interval). We then formulate the forecasting problem across point and probabilistic settings, distinguishing single- and multi-step strategies and showing how horizon choice shifts the balance between persistence, cloud-driven dynamics, and NWP-driven uncertainty. We propose a taxonomy covering (i) physics-based approaches (clear-sky and PV performance models), (ii) statistical and time-series baselines, (iii) ML/DL families — recurrent, convolutional, attention-based, and transformer architectures — and (iv) hybrid and ensemble methods, with attention to calibration and uncertainty quantification. Finally, we identify open challenges limiting robust deployment — non-stationarity, cross-site/climate generalization, probabilistic calibration, and explainability — and highlight emerging trends such as self-supervised learning, foundation models, and digital-twin-inspired forecasting loops.